Labs

Sakana AI welcomes Jürgen Schmidhuber as Chief Scientific Advisor. https://sakana.ai/schmidhuber/ Sakana AI is incredibly proud to announce that Jürgen Schmidhuber, universally recognized as the father of modern AI, is officially joining Sakana AI as Chief Scientific Advisor. For nearly four decades, Jürgen has explored how machines can learn to learn. His foundational work in the 1990s drove core advancements in deep learning and established early frameworks for world models. Crucially, his pioneering innovations in meta-learning opened the very path toward recursive self-improvement. These ideas have already shaped our own research, from the Darwin Gödel Machine to The AI Scientist. Now Jürgen will help guide our newly formed RSI Lab, whose objective is to trigger a compounding cycle of scientific discovery aimed at improving machine intelligence. We are assembling a critical mass of world-class experts in Tokyo to make this a reality. Welcome, @SchmidhuberAI !

English is untouched: pooled score 60.3 → 60.1. Language consistency is cheap but reliable termination is the open problem. Data came from ~800k traces made by prefilling a teacher's reasoning with German openers. Full post here: https://aleph-alpha.com/en/blog/through-the-valley-of-tears-cold-starting-german-reasoning-in-llms/

The loss is not wrong answers. It's traces that never finish. At the bottom of the valley, 1 in 4 German traces loops until the context limit. The loop rate vs score across German benchmarks is ρ = −0.80 to −0.89. Climbing out takes the right data, not more data. Total German share predicts nothing (|ρ| ≤ 0.34), while domain-matched share does: German math → German AIME ρ = +0.86, German chat → German IFEval +0.66. At ×16 math, AIME is back to 67.3. Loops persist at ~15%.

Setup: 13 SFT mixes on the same 3B-active MoE, ~16B tokens each. Only the German share varies, ×0 to ×16 per data group (chat, math, tools). Eval on AIME'26, IFEval, MuSiQue in EN and DE, plus a native German culture benchmark. The baseline with zero German data wins every German reasoning benchmark by thinking in English: 70.2 on German AIME. Add any German reasoning data and the model switches to German thinking in 100% of traces. Score: 48.3.

Reasoning models think in English, even on German prompts. We asked what it costs to make one think in German. Answer: a valley. Small doses of German reasoning data hurt, large doses mostly recover. https://x.com/Aleph__Alpha/status/2103139585039511700/photo/1

Time to catch up with Apollo Go in London. 🇬🇧 More than a month into testing with @lyft, our Apollo Go RT6 vehicles have been getting to know the capital's complex streets and traffic patterns. Take a look at how testing is coming along. You might recognize a street or two. ↓ https://x.com/Baidu_Inc/status/2103122281262198882/video/1

live on GMI Cloud 🔥 https://twitter.com/gmi_cloud/status/2102564626726854803

🚀 Tencent Hy Translation just landed.Powered by Hy-MT2. 33 languages. 5 Chinese minority languages & dialects. Voice. Photo. Full offline — on-device, no network required. Already live in 12 countries and regions. Travel, drive, work or read abroad. Accurate. Natural. Always available. ⏬ ⏬⏬ https://apps.apple.com/us/app/tencent-hy-translate/id6801428074

We put Muse Realtime Avatar head-to-head with two leading commercial avatar systems in their own live-call products. Raters held 2–3 minute conversations with matched avatar identities, then compared visual quality, sync, character consistency, and mannerisms. Muse Realtime Avatar came out ahead on overall preference. Read our research blog to learn more about Muse Realtime Avatar: https://research.meta.ai/blog/bringing-your-muse-to-life

Muse Realtime Voice and Muse Realtime Avatar form a unified streaming architecture connecting conversational intelligence, voice, and video embodiment via a shared speech-token stream. Muse Realtime Voice generates speech tokens encoding both content and prosody. Muse Realtime Avatar consumes this shared stream and generates streaming video. By using a fixed-length history as motion context for subsequent chunks, the system maintains bounded computation across arbitrary conversation lengths while generating synchronized voice, lip motion, and expressions. Live video streaming needs to respond instantly while remaining visually consistent over long conversations. We achieved this by distilling a large 40-step diffusion teacher with 3-way CFG (120 evaluations per video chunk) into an unguided 2-step causal student with a fixed-length KV cache. Self-forcing helps the student resist drift and maintain near teacher quality with 60x fewer evaluations.

More ways to build with MiniMax-H3. 🐮 Great to see model optimization and AMD inference engineering come together to give creators a faster feedback loop.⚡️ https://twitter.com/NunchuxAI/status/2102797754724757552

⚡️ As LLM reinforcement learning scales to larger GPU clusters and more training data, training efficiency becomes a first-order concern. Our new research revisits classical critical-batch-size theory and extends it to online LLM RL, where the model generates its own training data and rollout generation and training scale differently. Across GRPO and PPO, we find that learning-rate retuning can preserve learning per response over a bounded range of batch sizes. On fixed hardware, scaling up the batch size improves PPO generation-stage throughput by up to 2.29×, while our best measured GRPO configuration reaches the same validation target in 29% less time. 🚀 Read the full research: https://hy.tencent.ai/research/100116

In the Democratic Republic of the Congo, global health organizations including @CEPIvaccines, @WHOAFRO, and @inrb_kinshasa are using Claude to accelerate their response to an outbreak of an unusual Ebola variant. Read the full piece here: https://www.anthropic.com/features/ebola-response

We’re demonstrating how frontier models have continued to improve in realistic mental health conversations with MentalHealthBench. This new open benchmark was built with input from more than 80 mental health clinicians. We’re releasing it openly so other researchers can examine the methods, run their own evaluations, and build on the work. https://openai.com/index/introducing-mentalhealthbench/ Most mental health benchmarks focus on emergency situations. MentalHealthBench is designed to cover the full spectrum of mental health conversations that people bring to AI - from everyday support to more acute crisis scenarios. https://x.com/OpenAI/status/2102837575568519450/photo/1

Today at #ClimateWeekNYC, Ai2 and @GlobalFishWatch announced a partnership to shape how AI and AI agents enter ocean monitoring and enforcement. Purpose-built. Open. Accountable to the people who use them. https://skylight.global/news/ai2-gfw https://x.com/allen_ai/status/2102829696840839630/photo/1

This is the first result from our new molecular biology lab, where a team of Anthropic biologists is using Claude to explore and accelerate fundamental biology research. There, Claude works through data and literature to generate hypotheses and candidate biological systems to study. After our scientists review Claude’s hypotheses, they test the most promising ideas, with all lab work done by our scientists. We’d like to extend this approach to a broad range of problems—in genomics and in other fields. If you have a proposal for a research question, we’d like to hear from you.

Claude has discovered a previously unknown enzyme system hidden in the DNA of bacteriophages. Beside the enzyme’s gene sits a long array of repeating DNA—a structure that looks somewhat similar to CRISPR. We don’t yet understand what this system does, but only a handful of known systems share its features, and all of them are able to cut, copy, and paste DNA. Historically, the discovery of such programmable systems has helped revolutionize medicine. CRISPR, for instance, is now the foundation of genetic medicines. But it will take much more work to learn what this system does, and whether it can be put to similar use. Read more: https://www.anthropic.com/news/claude-discovers-novel-enzyme-system

Set up your first Vault in minutes: https://cohere.com/solutions/model-vault

Model Vault is now available in Canada 🇨🇦 For those in the "true North", you can now take full control of your AI with auto-scaled workloads, single tenancy, and lower TCO. Cohere models. Completely private. https://x.com/cohere/status/2102823381565382965/photo/1

We heard you loud and clear. ChatGPT Voice can now: - Use plugins like your email, calendar, and Slack. - Be powered by GPT-6 Astra, Sol, and Luna. - Be used in ChatGPT Work on web and mobile, so you can create docs, decks, sites, and spreadsheets or tackle complex tasks in the browser, just by talking. Rolling out globally today in the latest version of the app.

Rolling out starting today: — Developers: Both models in @GoogleAIStudio and the Gemini API — Consumers: Gemini 3.8 Flash TTS in @Gemini_Notebook and Gemini 3.8 Flash-Lite TTS in Google Vids — Coming soon: Both models in Gemini Enterprise https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-text-to-speech/

We’re launching Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS ⚡️ Our most expressive audio models yet let you create custom voices across 100+ languages or pick from 2,000+ ready-to-use ones. You can direct back-and-forth conversations, guide the delivery line-by-line, and add natural cues like <laughs> or an active listening interjection like |mhm| all while generating hours of consistent, glitch-free audio. Sounds pretty cool, right? So… how should you use them? — Gemini 3.8 Flash TTS: Need to design bespoke vocal personas from scratch and with line-by-line level control? This is the model! Built for high-fidelity creative production like gaming, immersive audiobooks, and podcasts. — Gemini 3.8 Flash-Lite TTS: Want the AI to automatically adjust its tone and pacing on the fly for near real-time voice agents? This is your engine! Built for cost-efficient scale, high-volume dubbing, and bulk audio creation.

Create and deploy custom audio with our new text-to-speech models: 🔵 Gemini 3.8 Flash TTS: Design unique voices with distinct accents and characteristics. 🔵 Gemini 3.8 Flash-Lite TTS: Built for efficiency and scale, choose from your created styles or our expansive production-ready library. Fine-tune the delivery line by line, shaping pacing, emotion, and cues like laughs or pauses. All generated audio is watermarked with SynthID so it can be reliably identified as AI-generated. Start building with the Gemini API via @GoogleAIStudio. Find out more → https://goo.gle/4dw40ns

Benchmark suite: https://x.com/Alibaba_Qwen/status/2102727421128810774/photo/1

Performance of Mobile-Use Agent: https://x.com/Alibaba_Qwen/status/2102727414317199671/photo/1 Performance of Mobile Creative Agent: https://x.com/Alibaba_Qwen/status/2102727417781772404/photo/1

Performance of Mobile Planner Agent: https://x.com/Alibaba_Qwen/status/2102727410093601172/photo/1

Introducing Qwen Intelligence, bringing personal intelligence within everyone's reach. 📱✨ It launches with three SOTA agents: 🥳 - Mobile Planner Agent: plans, decomposes & orchestrates complex tasks. #1 on MobilePA-Bench, MobilePA-Bench Business & Memory. - Mobile-Use Agent: gets things done, API-first with GUI fallback. MobileWorld 82.1, MobileWorld-Real 92.2, AndroidDaily 97.2, 90% end-to-end success rate. - Mobile Creative Agent: turns one sentence into ready-to-use creations. Image generated in 3s, about 2x faster than leading peers. We're also opening up our benchmark suite: MobilePA-Bench, MobileWorld, MobileWorld-Real, and MobileWorld-Safety, covering planning, cross-app execution, real-device performance and safety. 🔗 Learn more about the agents: - Qwen Intelligence official website: https://www.qwenintelligence.com - Mobile Planner Agent: https://github.com/Tongyi-MAI/Qwen-Planner-Agent - Mobile-Use Agent: https://tongyi-mai.github.io/Qwen-UI-Agent/ - Mobile Creative Agent: https://arxiv.org/abs/2608.16887 🔗 Explore our open benchmark suite: - MobilePA-Bench: https://tongyi-mai.github.io/MobilePA-Bench/ - MobileWorld (GitHub): https://github.com/Tongyi-MAI/MobileWorld - Leaderboard: https://tongyi-mai.github.io/MobileWorld/#leaderboard

⚡ Meet Qwen-Audio-3.1! ASR, TTS & Realtime are fully upgraded, joined by two new models: TTS-Next for audio creation and ASR-Next for audio understanding. Five models, one complete audio stack: understanding, generation, interaction & creation. Plus big price cuts across the lineup: TTS ~70% off, Realtime ~85% off, and ASR up to 95% off. Highlights: 🥳 - ASR: stronger multilingual & dialect recognition, plus native polishing that auto-removes fillers & repetitions for cleaner, more logical transcripts. - ASR-Next: supports multi-speaker ASR with speaker labels, timestamps & aligned transcripts, and understands emotions, ambient & machine sounds for sound captioning, event localization, audio QA & reasoning. - TTS: multilingual & dialect synthesis with natural cross-lingual voice transfer; control emotion, speed & style via simple instructions. - TTS-Next: unified LM + diffusion framework generating voice, sound effects & background audio in one pass for audiobooks, podcasts, games & ads. - Realtime: speak & listen at once with anytime interruption, just like a real call; it even slows down and responds empathetically when it senses a low mood. Unlock the full potential of Qwen-Audio-3.1! 👇 - Blog: https://fun-resource-shanghai.oss-cn-shanghai.aliyuncs.com/cuijiayan.cjy/tmp/exp/qwen_audio_3_tts_blog_review_260918/index.shtml?Expires=2105366399&OSSAccessKeyId=LTAI5tQrCBwj82sVMCWoSmzE&Signature=9ZaZIEahoN41Zb9MFJjf34G%2BSCM%3D - Qwen-Audio-3.1-ASR: https://www.qwencloud.com/models/qwen-audio-3.1-asr-flash-filetrans - Qwen-Audio-3.1-Realtime: https://www.qwencloud.com/models/qwen-audio-3.1-realtime-plus - More APIs: coming soon @qwen_cloud

Free for two weeks！Hy Image3.5 preview is live on OnSolo. Short drama character sheets. Full-motion video game assets. Keyframes. Characters hold across every episode. Edits refine instead of restarting. https://twitter.com/OnSoloAI/status/2102649675195228354

ComfyUI ✖️ Hy Image3.5 preview https://twitter.com/ComfyUI/status/2102356846585786786

Thanks @arena for the recognition! 🏆 Qwen-Image-2.1 is now the #1 open-source model in both the Image Edit and Text-to-Image Arenas. Try it now and show us what you create! 🎨 https://twitter.com/arena/status/2102416020678008986

GPT‑6 Sol and Luna build on Astra’s advances in alignment, showing improvements over their GPT-5.6 counterparts. https://x.com/OpenAI/status/2102460992202420225/photo/1 GPT-6 Sol and Luna roll out today in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users. Both are also available in the API. Free and Go users can try GPT-6 Luna in the desktop app. https://openai.com/index/introducing-gpt-6-sol-and-luna

Higher usage limits and lower cost give you more flexibility and room to iterate. https://x.com/OpenAI/status/2102460984782938221/photo/1

Please welcome GPT-6 Sol and GPT-6 Luna to the GPT-6 universe. GPT-6 Sol and Luna build on the advances behind GPT-6 Astra, bringing much of its strengths into faster and more affordable models to support work at scale. We’ve also made caching and inference more efficient, and we’re passing the savings directly to you: 50% lower API prices for Sol and Luna compared with GPT‑5.6 promotional pricing.

Huge thanks to @diyerxx and the team for this work 🫡

We’ve open-sourced onPanda 🐼 — the tool we use internally for LLM data annotation and model inspection. The workflow is simple: find an error, correct the token, and let the model continue. ✍️ Data annotation - 52% lower median annotation time vs. manual post-editing - SFT + preference data in one workflow, with high on-policy fidelity (ΔPPL <1% vs. the model’s resampling baseline) - Precise token-level supervision with paired positive/negative examples, plus agent-trajectory annotation across image, audio, and video 🔎 Model inspection and debugging - Inspect token probabilities and top-k alternatives, steer decoding token by token, and explore SVG generation, web development, and agent tasks directly in the browser. Try it (mobile-friendly): https://onpanda.diyer22.com Paper: https://huggingface.co/papers/2609.24983

As part of our efforts to pace the frontier, we’re committed to supporting independent assessments with deep levels of access across training, evaluation, and deployment. That access should enable third party assessors to challenge our assumptions, identify risks we may have missed, and reach their own conclusions about the effectiveness of our safeguards. We’re outlining four priority areas for deeper assessment, alongside principles for rigorous, secure, and independent work: https://openai.com/index/priorities-principles-third-party-assessments/

Claude Opus 5.5 is available today. https://twitter.com/claudeai/status/2102435511222890900

Step Code also includes StepPage: publish a local static site to a shareable URL with one command. Version management and rollback are built in, so development, debugging, and delivery can stay in the same terminal. Install Step Code on macOS, Linux, or WSL: curl -fsSL https://static-openapi.stepfun.com/stepcode/install.sh | bash Windows: WSL is recommended. PowerShell support is currently in beta. Issues and pull requests are welcome: https://github.com/stepfun-ai/Step-Code

Step Code is built for strong task completion without wasteful token use. On Terminal-Bench 2.1, it passed 72/89 tasks (80.9%), tying for the highest pass rate among the harnesses in our evaluation while using fewer tokens than the other tied leaders. On Multi-Frame, it passed 110/150 tasks (73.3%) and averaged 5.09M tokens/task—the highest pass rate and lowest token use among six harnesses evaluated.

Introducing Step Code v0.1.0. Swift execution. High token efficiency. Long-horizon reliability. Now open source under the MIT License. Step Code handles the full development loop—from reading and editing code to running tests and shipping—from one CLI. - 80.9% on Terminal-Bench 2.1 in evaluation - 73.3% on Multi-Frame, our 150-task long-horizon benchmark - One-command static site publishing with StepPage GitHub: https://github.com/stepfun-ai/Step-Code

One move ahead at ALL-IN ♟️ We hosted Grandmaster @MagnusCarlsen for a night all about AI, technology, and of course chess. https://x.com/cohere/status/2102399340215996726/photo/1

🎨 Hy Image3.5 preview is live in Miora. Edits that keep what already works — same canvas, your brand rules already remembered. Free for two weeks. https://twitter.com/Miora_Design/status/2102216310545608765

Meet the new Kimi Browser Extension, formerly Kimi WebBridge. From your browser sidebar, you can chat with Kimi to navigate websites, fill out forms, and get things done. For repetitive tasks, record your steps once and save them as a skill. Kimi can take it from there next time. Available now on http://kimi.ai/products/kimi-browser-extension and the Chrome Web Store.

Two teams took home the 2026 Baidu Highest Award at our annual party yesterday for their work in AI infra and AI adoption. 🎉 Belief in the value of technology is our true edge, and the people behind it deserve a celebration. Cue the AI demos, live music and good food. Here's to building together!

Share some genius cases generated by Hy Image3.5 preview. 😎 https://x.com/TencentHunyuan/status/2102324577976426830/photo/1 https://twitter.com/TencentHunyuan/status/2102226552310419473

The AI says the site is done. Then the homepage errors, the buttons overlap, and it added a login flow you never asked for. 🙃 Introduces WebCraftBench: agents actually use the live app, coverage-guided exploration finds what never got reached, then we score aesthetics, usability, and whether the original request was met. On 197 human-validated pairs, it matches human preference 85.3% of the time. Paper: https://arxiv.org/abs/2609.15387

Day-0 OpenVINO support from @inteldevs! 🥳 Qwen-Image-2.1 is ready to run optimized on Intel hardware. One open-weight checkpoint for both generation and editing. 👇 https://twitter.com/inteldevs/status/2102110058360209902

Kimi K3 is now on Amazon Bedrock! Run coding, document analysis, and extended agent workflows with Bedrock's access, encryption, and auditing controls. Explicit prompt caching supported. Start building with K3 on AWS 👉 https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-moonshot-ai-kimi-k3.html https://x.com/Kimi_Moonshot/status/2102244258531213596/photo/1

Step 5 Preview pushes our intelligence–cost Pareto frontier outward. It comes in at 44 on the Artificial Analysis Intelligence Index, $0.71 per task. Thanks @ArtificialAnlys for putting Step 5 Preview through the full evaluation. More to come on Oct 15. https://twitter.com/artificialanlys/status/2102213621963243704

Hy Image3.5 preview is live. 🚀 Professional-grade image generation, +30% win rate in human eval vs Hy Image3.0 Both Text to image & Image to image available. Up to 2K. Better Consistency. API: https://console.tencentcloud.com/tokenhub/models/detail?modelId=hy-image-v3.5-preview Priced for everyone. $0.024 per image on Tencent Cloud API. We only charge for what we generate — your reference images are free. Two weeks free (Only in OnSolo and Miora): Miora https://miora.design/ OnSolo https://onsolo.ai/ Try it and tell us where it breaks.

Start building with MiMo-V2.6. 🚀 MiMo Desktop and membership plans launch alongside Pro and Flash. Use both models with a subscription, or bring your own API key. MiMo-V2.6-Pro UltraSpeed is now available in Desktop and through the API, delivering up to 20× faster generation for workflows where response time matters. API pricing remains unchanged from V2.5. 💻 Desktop: https://mimo-ai.xiaomimimo.com/desktop/ 🔗 API: https://platform.xiaomimimo.com/ 🤗HF: https://huggingface.co/collections/XiaomiMiMo/mimo-v26

We’re open-sourcing Pro and Flash, MiMo-V2.6-Distill-Qwen-9B, the technical report, 7K+ RL task environments, an end-to-end RL framework and composable mini-harnesses. Reproduce, verify and build on the work. https://x.com/XiaomiMiMo/status/2102138582324625780/photo/1

MiMo has its own taste. 🎨 MiMo-V2.6 brings code, design and tool use together across interfaces, slides, SVGs, video and music. 🔹 Build frontend interfaces and presentations with coordinated typography, layouts, interactions and animation 🔹 Work with Figma and image/video generation tools to create visual assets 🔹 Produce videos from concept to final cut, combining motion, music and narration with MiMo-V2.5-TTS 🔹 Compose demo-level music, including an orchestral piece for around ten instruments, and convert the score to MIDI On Design Arena, Pro performs at a level comparable to Claude Opus 5 and GPT-5.6 Sol.

A co-pilot for scientific research. 🔬 Without RL tailored specifically to scientific research, MiMo-V2.6-Pro is showing promise in materials design and mathematical formalization. 🧪 Materials research Working with Xiaomi’s materials researchers, it proposed MOF materials for capturing PFAS “forever chemicals,” reviewed literature and patents, and ran computational screening to identify candidates for wet-lab validation. 📐 Formal mathematics It helped researchers formalize the full main theorem of Li–Yorke’s “Period Three Implies Chaos” in Lean 4. After revision and integration, the project spans 6,000+ lines of code, verified by Lean’s kernel with no unfinished proof placeholders.

From Vibe Coding to Vibe World. 🌍 MiMo-V2.6 brings together 3D spatial reasoning, multimodal perception and computer use to build and interact with richer environments. 🔹 Turn text, images or video into playable 3D worlds, coordinating agents to build scenes, write interaction logic and refine the results 🔹 Create Blender objects and scenes for animation, 3D printing and games 🔹 Control a Franka Panda arm in simulation through visual feedback 🔹 Use desktop tools to search, edit and process data — then inspect the results and adjust its next actions Build. Observe. Refine.

More intelligence. Same price. 📈 MiMo-V2.6 pushes the intelligence–cost Pareto frontier outward once again. 🔹 API pricing unchanged from V2.5, for both Pro and Flash 🔹 MiMo-V2.6-Pro sets a new price-performance record among Chinese models 🔹 At comparable intelligence, Pro costs just 1/20 to 1/60 as much as leading international models

Introducing Xiaomi MiMo-V2.6 — Pro & Flash. Frontier intelligence, all the modalities, built in public. 🔹 Two omnimodal models, advancing through scaled reinforcement learning 🔹 Pro performs on par with Claude Opus 5 and GPT-5.6 Sol across most agent benchmarks 🔹 Pro scores 46 on the Artificial Analysis Intelligence Index — the highest among open-source models 🔹 Stronger coding, computer use, 3D reasoning and creative capabilities 🔹 Open model weights, technical report, RL environments and training code Blog：https://mimo.xiaomi.com/mimo-v2-6

Before you meet MiMo-V2.6, step into its world. Every 3D model, game, slides, video, music and image in MiMo Gallery was generated by MiMo-V2.6. Stay for the piano. Play a game. Catch the fireworks. A first look at what's coming: https://mimo.xiaomi.com/mimo-gallery/ https://x.com/XiaomiMiMo/status/2102093582891147339/photo/1

We’re working with an independent advisory group of mathematicians to help OpenAI responsibly share advances in AI and mathematics. The group will advise on how we assess and communicate new mathematical results, uphold academic and professional standards, and build tools that support mathematical research and learning. Through this work, we want mathematicians to be at the center of shaping how AI supports mathematical understanding and how its benefits reach the wider community. https://openai.com/index/advisory-group-on-mathematics-and-ai/

https://kooko.ai/

Kooko means business. Bring on the big briefs! Formerly Oreate AI, the all-in-one AI workspace has reached 10M+ international users across nearly 200 countries and regions in its first year. With Kooko, you can: - Research, analyze data, and create content using the files and sources you give it access to - Turn a brief into documents, presentations, spreadsheets, and more, then refine them in the built-in Office editor - Build on previous projects with memory that saves your edits, preferences, and completed work as context for future tasks With apps for desktop and mobile, Kooko is ready to take on your next assignment, wherever you are. Put Kooko to work👇

Thanks @sgl_project for the day-0 support! 🙌 SGLang-Diffusion now serves Qwen-Image-2.1: text-to-image generation, multi-image editing, and transparent RGBA output. Try it out! 🎨 https://twitter.com/sgl_project/status/2101697330420589054

Qwen-Image-2.1 × @HuggingApps: live demo on Spaces! 🖼 One single checkpoint for generation and editing. Try it in your browser, no setup needed. 👇 https://twitter.com/HuggingApps/status/2101708412148990429

Qwen-Image-2.1 is now supported in ComfyUI! Try it now and share your creations! 🖼️ https://twitter.com/ComfyUI/status/2101662316051997092

Thanks @vllm_project for the day-0 support! 🙌 Generate, edit, transparent output — one model, ready to serve. Details 👇 https://twitter.com/vllm_project/status/2101665920565629318

Qwen-Image-2.1 also improves visual quality over its predecessor, particularly in the aesthetics of text and portraits. Below are several examples of text rendering. https://x.com/Alibaba_Qwen/status/2101659324980625533/photo/1

Qwen-Image-2.1 supports several ways to specify local edits. The example below uses circles to identify three regions and asks the model to “remove the metal watch in the blue circle, change the hair in the red circle to black, and replace the area in the green circle with gray short-sleeved linen pajamas,” performing removals and modifications in all three regions at once. Qwen-Image-2.1 supports a variety of editing tasks while balancing performance across them. For example, given a three-view character reference, the model generates a complete storyboard. https://x.com/Alibaba_Qwen/status/2101659321549660610/photo/1

Qwen-Image-2.1 supports up to 10 reference images and can produce high-quality results. https://x.com/Alibaba_Qwen/status/2101659314180280769/photo/1

Qwen-Image-2.1 now supports transparent image generation natively, and naturally supports editing transparent images as well. https://x.com/Alibaba_Qwen/status/2101659310598361194/photo/1

Performance: https://x.com/Alibaba_Qwen/status/2101659306341097575/photo/1

Meet Qwen-Image-2.1, the most balanced and cost-effective image generation model in the Qwen-Image series! Now open weights! 🎨 A unified model for both generation and editing, delivering top-tier quality in a lightweight package. Highlights: 👀 - Compact & exceptionally fast: A lightweight 7B architecture that outperforms most closed-source models, with drastically accelerated inference for multi-image inputs. - Native transparency: Natively generates and edits RGBA layers, enabling seamless compositing and text editing within transparent images. - Versatile, high-fidelity editing: Supports up to 10 reference images and precise local control while preserving strict fidelity for portraits and products. - Broad coverage & stunning aesthetics: Excels at panoramas, infographics, and virtual try-ons, delivering realistic textures and elegant typography. Start to create your next masterpiece with Qwen-Image-2.1! 🖼️ - Blog: https://qwen.ai/blog?id=qwen-image-2.1 - GitHub: https://github.com/QwenLM/Qwen-Image-2.1 - Model Scope: https://www.modelscope.cn/models/Qwen/Qwen-Image-2.1 - Hugging Face: https://huggingface.co/Qwen/Qwen-Image-2.1

Try Step 5 Preview: https://platform.stepfun.ai Docs: https://platform.stepfun.ai/docs/en/guides/models/step-5-preview Join our Discord: https://discord.gg/F4EAbdDzRu

Finance is a particular focus for Step 5 Preview. Financial work has to stand up to scrutiny. We evaluate Step 5 Preview on its ability to identify and verify reliable information, reconcile differences across reports, make assumptions explicit, and produce internally consistent forecasts and reproducible valuations. FinStepBench tests these capabilities across LiveSearch, CorporateValuation, and DeepResearch. We also evaluate Step 5 Preview on FrontierFinance across six investment use cases.

Step 5 Preview is built for professional knowledge work — from large-scale research to structured analysis and interactive reporting. It coordinates research at scale and turns evidence into finished, auditable deliverables. In one agent action, it coordinated 950 web fetches and assembled 300,000 monthly records across 1,000 locations over 25 years; in another, it produced a 17-sheet analytical workbook with source reconciliation, formulas, and trend models. Its outputs span technical engineering, creative production, analytical reporting, and public communication.

Step 5 Preview works across software environments and sustains execution over long horizons. Its capabilities extend from software engineering and web applications to 3D workflows and programmable hardware. Over longer horizons, Step 5 Preview keeps track of prior results, uses execution feedback to decide what to try next, and continues iterating. We test this behavior in runs lasting up to 24 hours, including tasks involving GPU kernel optimization and automated post-training.

Introducing Step 5 Preview: Advancing the Pareto Frontier. Step 5 Preview is our new flagship model for agentic work, delivering frontier-level performance across software engineering and professional knowledge work, with particular strength in finance. - 600B total / 27B active MoE, with 1M context + Vision - Substantially lower task cost at comparable intelligence - Broad software engineering capabilities with sustained execution over long horizons Try Step 5 Preview: https://platform.stepfun.ai Model page: https://www.stepfun.com/step-5-preview Open weights on Oct 15.

All out for ALL-IN ⚡ @AstonMartinF1 and @MagnusCarlsen were in attendance with Cohere for Canada’s biggest AI conference https://x.com/cohere/status/2101412198564131149/photo/1

A great example of practical AI solving real-world problems! Faster than ever. Thanks for building with Qwen. @cerebras Try it out for yourself! 🏠 https://twitter.com/cerebras/status/2100997025773023403

Cases and more details: https://x.com/Alibaba_Qwen/status/2101206711461966067/video/1

Performance: https://x.com/Alibaba_Qwen/status/2101206708202987709/photo/1

Meet Qwen3.8-LiveTranslate, Qwen's next-generation real-time simultaneous interpretation model! 📢 Built on an Interleave architecture, it improves faithfulness, fluency, and conciseness while reducing average lagging (LAAL) from 2.8s to 2.3s across 60 languages. New capabilities: 🙌 - Real-time speaker diarization — distinguishes speakers in multi-party speech and preserves each speaker's voice through more stable voice cloning. - Synchronized bilingual display — source and translation on screen together. - Long-context disambiguation — leverages conversation history to clarify names and terminology for consistent translations. Let's try Qwen3.8-LiveTranslate! 🥳 - Blog: https://qwen.ai/blog?id=qwen3.8-livetranslate - QwenCloud: https://www.qwencloud.com/models/qwen3.8-livetranslate-flash-realtime

We’re partnering with Accenture on independent evaluation of frontier AI—part of our recent commitment to embed evaluators at Anthropic. Both we and Accenture expect to invest at least $1 billion to build capacity in this area over the next five years. https://www.anthropic.com/news/accenture-embedded-evaluation

It’s time for our end-of-week recap 👇 — Gemini 3.8 Live and 3.8 Live Extended Thinking, our most advanced live dialogue audio models yet — Dreambeans, an experiment from @GoogleLabs that curates a daily personalized collection of stories, is now GA — CC from @GoogleLabs has expanded from a personal productivity tool into a shared agent, designed to help families and households coordinate logistics, schedules, and daily tasks — Google Pics, a new @GoogleWorkspace tool that lets you generate, refine, and co-create images, is now GA — AlphaGenome Atlas, @GoogleDeepMind's new interactive platform for genomics discovery

CONVERGENCE, in review: We launched AI for Empowerment at Berlin Art Week with one goal: spark the conversation around how technology should empower you to live the life you desire. https://x.com/cohere/status/2101000146456571906/video/1

MiniMax Code CLI is open now: https://github.com/MiniMax-AI/minimax-code https://twitter.com/MiniMaxAgent/status/2100928718541853038

We are aware of a claim alleging unauthorized access to our systems. Following a thorough investigation, we have found no evidence to support this claim and can confirm that our systems have not been compromised.

We're proud to contribute as an AI partner to the newly launched @Singtel AI Pass, supporting Singapore's SkillsFuture AI Subscription initiative. 🇸🇬 MiniMax H3 (@Hailuo_AI), MiniMax Agent (@MiniMaxAgent) and MiniMax Audio are all included, giving eligible learners across 200+ SWDA-supported AI courses hands-on access to premium AI tools and the opportunity to build practical, real-world AI skills. This partnership brings our mission, "Intelligence with Everyone", to life, empowering learners across Singapore to explore, create, and build confidently with AI. Learn more: https://www.singtel.com/personal/products-services/lifestyle-services/ai-pass

Three paths to faster video attention: compute the same interactions more efficiently, compute fewer in full, or change how information is mixed. Here’s a visual guide. 👇 Thanks to Nunchux AI and collaborators for VC-Attention, bringing training-free low-bit acceleration to MiniMax-H3, with better fidelity than SageAttention2 in the B200 evaluation. The approach balances speed and fidelity: V-Smooth reduces value quantization error, while ExpCast-FP8 makes softmax faster through approximation. Excited to see the community keep building on H3. Could combining low-bit computation with sparse methods like Sol-Attn push efficiency further? We’re looking forward to seeing that explored.

Cases and more details: https://x.com/Alibaba_Qwen/status/2100785971239534679/video/1

Performance and price: https://x.com/Alibaba_Qwen/status/2100785967309496512/photo/1

🚀 Meet Qwen3.8-Omni-Flash, Qwen's first omni-modal model built around agentic capabilities! Native audio-video understanding, reasoning, and tool use come together in one model: understand the content, plan the task, execute with tools, and deliver the result. Highlights: 🥳 - Audio-video intelligence that gets things done: jointly reason over what's seen and heard, and orchestrate tools across long workflows to auto-edit vlogs, translate short videos, and turn movies into recaps. - A major leap: approaching Gemini 3.8 Flash in audio-video capabilities; +19.5 points on average in agent performance across WildClawBench-MM & UniClawBench. - 1M-token context with agentic perception: actively explore long videos and locate key moments with higher accuracy, using 51.8% fewer tokens than static understanding on OmniVideoBench. Video input costs are reduced by about 89% compared with Qwen3.5-Omni-Plus, making long-form audio-video understanding and agentic workflows more affordable than ever. To help you build apps around Omni, we're also open-sourcing Qwen-MM-Plugins and Qwen-Live Harness! 🛠️ We can't wait to see what you build with Qwen3.8-Omni-Flash! 👀 - Blog: https://qwen.ai/blog?id=qwen3.8-omni-flash - Qwencloud: https://www.qwencloud.com/models/qwen3.8-omni-flash - Qwen Studio: https://chat.qwen.ai/ - API: https://www.alibabacloud.com/help/en/model-studio/qwen-omni - Qwen-MM-Plugins: https://github.com/QwenLM/Qwen-MM-Plugins - Qwen-Live Harness: coming soon https://github.com/QwenLM/Qwen-Live-Harness

次のパラダイムを探索する、Sakana AIの研究グループ「Frontier Intelligence Group」 https://sakana.ai/frontier-intelligence-group/#Japanese Sakana AIには、現在の主流とは異なる方向からAIの可能性を探る研究グループ「Frontier Intelligence Group（FIG）」があります。FIGは、知能はまだ「解決」しておらず、現在のAIと生物 の知能のあいだにあるデータ効率やエネルギー効率、汎化能力の差を、研究として正面から扱うべきだという立場から出発しています。 本記事では、FIGの考え方と、Continuous Thought Machines、予測符号化による学習、スパースなTransformer、AI Picbreeder、Smart Cellular Bricks などこれまでの研究を紹介しています。

Introducing the Sakana AI Frontier Intelligence Group 🪷 https://sakana.ai/frontier-intelligence-group Current AI systems are incredibly capable, but is intelligence “solved”? And if not, what’s missing? At Sakana AI’s Frontier Intelligence Group (FIG), we believe that there are still breakthroughs to be made in AI. The Transformer and language modeling may be incredibly powerful, but it doesn’t mean that better alternatives don’t exist. Natural intelligence still beats artificial intelligence across many dimensions. Agents lack the deep insights and creativity of humans. Individual models require far more data than the brain to learn robustly, and require far more energy to run. If we set these as targets, what kinds of AI systems could we develop? Research at FIG has sought to address the gaps between natural and artificial intelligence. Here are some of our works, and the fundamental research questions that motivated them: • Continuous Thought Machines: How can we improve information processing by leveraging temporal dynamics? • Augmented Lagrangian Predictive Coding: How can local learning solve multilayer credit assignment? • Sparser, Faster, Lighter Transformer Models: How can we massively increase data efficiency and generalization? • The AI Picbreeder Experiment: How can we make artificial open-ended systems? • Smart Cellular Bricks: How can physical systems achieve collective intelligence and self-repair without a central brain? We hope that this encourages other researchers to also explore different paradigms, and take a leap of faith with us. After all, in the words of a dear friend of ours, “greatness cannot be planned”.

To show what these optimizations make possible, we’re partnering with Adaptyv Bio on a protein design competition. Together, we’ll be experimentally validating over 5,000 designs. We're providing up to $1 million in Claude credits plus funding alongside Adaptyv for experimental validation. Modal is contributing up to $250,000 in compute and Twist Bioscience is providing DNA. Learn more on Adaptyv’s Proteinbase: https://proteinbase.com/competitions/anthropic-adaptyv-2026 And sign up for the competition here: https://docs.google.com/forms/d/e/1FAIpQLSc0Hz1ZWYTt_wkn76ViVxDghmEhG_OeVEcj9YGHxLWqxF1kWw/viewform You can find all of the code on GitHub: https://github.com/anthropics/uplifting-biomolecular-modeling And the full results in our technical report: https://www-cdn.anthropic.com/d8ca26d0d205708d26c7337cf4cfe7cb52e9b671.pdf

Biologists use specialized open-source models for tasks like modeling the structure of molecular systems, designing drug-like molecules, and predicting the effects of genetic mutations. But these models are often expensive to run, potentially limiting their impact. In our latest Science Blog, we share how Claude was able to optimize inference for more than 30 open-source models, making them 4x faster on average, partly by writing custom software for GPUs. We’re open sourcing all of the optimization code. Read more: https://www.anthropic.com/research/claude-uplifts-biomolecular-modeling

AI systems are getting more powerful, and they're increasingly being used to build the next version of themselves. We want to illuminate that progress for the public. Today, we're sharing three measurements that help track AI development: 1. How much AI R&D is done by AI. 2. How well AI agents are overseen. 3. How compute is allocated. We provide a snapshot of these metrics from inside Anthropic. Any frontier developer could publish the same measures, and third parties could verify them. As the world considers pacing the frontier, we should do everything possible to minimize the gap between what frontier labs know and what the public knows. This means better measuring the development of AI, publishing our findings, and giving society an opportunity to decide how to use this information. Read the full post and methodology: https://www.anthropic.com/institute/measuring-pace-of-ai-development

Astra for Law will initially be offered to selected firms through Trusted Access in ChatGPT and Codex, with API access coming soon. We’ll maintain the legal configuration so builders can focus on their own products and workflows. We’ll build the next chapter of Astra for Law alongside the lawyers and legal technology partners who put it into practice. https://openai.com/index/astra-for-law/

Astra for Law pairs GPT-6 Astra with instructions for legal analysis and writing, settings for thorough work, and a new Legal Search Index. The index searches U.S. case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs—with sources added daily. That helps lawyers move from the facts of a matter to relevant authorities and supporting passages they can examine themselves. The people who know the work should shape the tools. With our engineers, Sullivan & Cromwell built an agreement analyzer, Ropes & Gray built an M&A diligence system, and Cooley built GO Public for IPO preparation. Each brings the firm’s expertise into workflows its lawyers can review, challenge, and refine. We’re also expanding privacy and governance controls for eligible firms through Trusted Access. We’re also launching 26 partner-built plugins and 47 community plugins for legal work in ChatGPT. Partners including @thomsonreuters, @harvey, @WeAreLegora, and @imanageinc connect specialist tools and knowledge. Community plugins built by lawyers and legal engineers give firms skills they can adapt and extend. The plugins bring more ways to build with the tools, knowledge, and people firms already rely on.

Astra for Law: Frontier intelligence built for your practice. A new offering powered by GPT-6 Astra with tools, settings, and context to support the expertise and judgment of lawyers and legal technology firms. https://x.com/OpenAI/status/2100679992720142459/video/1

Can a fully open model help make AI more “prosocial” through a game? @SohamPadia built Steering Arena, where players try to elicit kind & respectful responses from Olmo 3. Surprisingly, strings like “Undert! AH :-) Rog Appl)” were highly effective. 👇 https://allenai.org/blog/olmo-arena https://x.com/allen_ai/status/2100679444042052063/photo/1

Introducing Fugu Max and Fugu Ultra v2: Orchestrating the Pareto Frontier https://sakana.ai/fugu-max-release/ https://x.com/SakanaAILabs/status/2100647118859886906/photo/1

Today we’re opening applications for the Life Sciences Verification Program. Through the LSVP, life science professionals can use our models—including, for the first time, Mythos—with a new set of safeguards designed to enable the full range of biology-related work. We designed these new safeguards to provide a better experience for biologists and more protection from risk of misuse. The program is launching in beta for teams of all kinds—from academic labs to startups, pharma companies, and more. We will continue to improve the program and expand access to individual Pro and Max plans over time. Learn more about these access grants and apply: https://www.anthropic.com/news/life-sciences-verification-program

🐙 Sakana Fugu Max is now on Sakana Chat 🐙 Free for all to use. Try it: https://chat.sakana.ai Blog: https://sakana.ai/chat-fugumax/#English 🐡 https://x.com/SakanaAILabs/status/2100592398602506439/video/1 https://twitter.com/SakanaAILabs/status/2098233826816205275

Every second, millions of genome switches dictate how our cells function and adapt. 🧬 Working with @ScienceStowers, Atlas mapped 2,500+ regulatory patterns uncovering the DNA sequences that act as biological switches and volume dials to turn gene activity up or down across hundreds of cell types. These discoveries are just the beginning. AlphaGenome Atlas is freely accessible to empower researchers everywhere to decode the genetic causes of disease. Explore the database → https://goo.gle/4heuCvn

Scientists at @BroadInstitute search for the underlying causes of unexplained rare disease cases, but the list of possible mutations is vast. Using Atlas, researchers pinpointed a critical mutation in the DNM1 gene, which was confirmed and validated in the lab. https://x.com/GoogleDeepMind/status/2100586763341222176/photo/1 In collaboration with the @UniofExeter, researchers used AlphaGenome Atlas to analyze data from 54,000+ @uk_biobank participants. They found: 📈 A 22%+ boost in detecting rare genetic signals 🎯 It uncovered new DNA variants influencing the abundance of PLA2G7, a protein linked to metabolic health.

Researchers at @BroadInstitute, @UniofExeter, and beyond are already using AlphaGenome Atlas to better identify potential disease-causing DNA variants and interpret their role. 🧵 https://x.com/GoogleDeepMind/status/2100586760036143330/photo/1

We welcome the HKSAR Government's continued support for autonomous driving, as outlined in the first Five-Year Plan and the 2026 Policy Address. Confident in its future in Hong Kong, Apollo Go plans to apply for the commercial operation. Building on the fully driverless trials, we look forward to advancing the city's vision for driverless, scaled-up and commercialized AV operations and helping position Hong Kong as a global benchmark for commercial autonomous driving in right-hand-drive markets.

Apollo Go is out and about in Hong Kong. 🚘 Fully driverless testing continues on the city's roads — come along and see Apollo Go in action. ↓ https://x.com/Baidu_Inc/status/2100582198483333262/video/1

Smart Cellular Bricks: Towards Collective Intelligence for the Physical World Our blog post: https://sakana.ai/smart-cellular-bricks Nature Communications paper: https://www.doi.org/10.1038/s41467-026-75166-7

The latest issue of Scientific American (@sciam) features our Smart Cellular Bricks research: simple cubes that collectively recognize their own shape and repair themselves without a central brain. https://www.scientificamerican.com/article/these-smart-bricks-know-what-object-they-make-up/ What if 200 simple blocks, with no central controller and no idea where they are, could figure out what object they have built, and tell you exactly where pieces are missing? That is what Smart Cellular Bricks do. Led by Sakana AI researcher Sebastian Risi and collaborators at IT University of Copenhagen and Autodesk, each cube runs the same small neural network and communicates only with its direct neighbors. From purely local exchanges, the collective converges on the correct global shape, locates damage, and can even guide its own repair. No module is in charge. No module knows its position. From simulation to 197 physical bricks with 100% accuracy, this is a step toward machines that build and heal themselves.

【採用情報】Sakana AIで「防衛ドメインエキスパート（Defense Domain Expert）」を募集🐟 https://apply.workable.com/sakana-ai/j/397C4C43B5 防衛省・自衛隊の業務と組織を熟知した立場から、Sakana AIのAI技術を「現場で実際に使える形」へ翻訳し、ユースケースの定義から提案・導入・運用定着までを一気通貫で牽引して いただきます。 🐟 このようなご経験をお持ちの方を募集 ・幕僚監部（統合／陸上／海上／航空）またはそれに準ずる機関での勤務経験 ・情報、運用、防衛計画、指揮通信システムのいずれかの実務経験と専門性 ・防衛省・自衛隊の組織構造、意思決定プロセス、調達の仕組みへの深い理解 自衛隊とスタートアップの間の「翻訳者」として動ける方、ぜひご応募ください🐟

【GTMチーム拡充・創業メンバー募集】　Sakana AIは、研究から生まれたプロダクトを日本と世界の企業・官公庁に届けるGTM組織を大規模に拡充します。 Sakana Chat、Namazu、Marlin、Fuguなど、既に世に出ているプロダクトを事業に拡大するフェーズで、営業もエンジニアも「創業メンバー」としてGTM の仕組みをゼロから作り上げます。 ▼募集ロール ・Product Sales & Account Executive ・Forward Deployed Engineer (GTM) 研究・プロダクト・現場が分断されない環境で、世界クラスのAIを事業に変える仕事をしませんか。 ▼詳細・応募 https://sakana.ai/careers/?category=gtm 🎏

We’re sharing how GLM-5.3 helped build and optimize the inference infrastructure serving GLM-5.3-Flash. The system went from its first successful run to production readiness in less than two weeks, with end-to-end throughput tripling relative to the initial baseline. The key was dense feedback: local correctness tests, execution traces, microbenchmarks, and end-to-end measurements that enabled targeted hypothesis testing rather than reliance on aggregate performance metrics alone. https://z.ai/blog/glm-built-its-inference-infrastructure

🐡 Sakana Chatがアップデート 🐡 8月にはモデルも新しくなった「Sakana Chat」がさらにパワーアップしました。 （1）オーケストレーターモデルを最新モデルに刷新 （2）メモリー機能を追加 を新たに実装しました。 Sakana Chatを試す：https://chat.sakana.ai/ Blog：https://sakana.ai/chat-fugumax/ 🐙 https://x.com/SakanaAILabs/status/2100381722147721724/photo/1

We're sharing our new framework for tracking, investigating, and disclosing instances of model misalignment at OpenAI. The framework sets criteria and timelines for public disclosure, including when we haven’t yet fully explained or mitigated the behavior. More complex cases may require longer investigation or coordination with third parties. We’ll prioritize examples that reveal new misalignment mechanisms, meaningful changes in known behavior, or findings that challenge assumptions about safety or mitigation. Alongside the framework, we’re publishing six reports on instances of misaligned behavior we’ve observed during the training or evaluation of our models in the last six months. This is a starting point. We’ll refine the process through experience and public feedback, and share more reports on an ongoing basis. https://openai.com/index/model-misalignment-reporting-framework/

Huggingface Space: > https://huggingface.co/spaces/stepfun-ai/StepAudio-3-Music API docs: > https://platform.stepfun.ai/docs/en/guides/models/stepaudio-3-music Thesis: > https://arxiv.org/abs/2609.1603

StepFun and ACE Studio present StepAudio 3 Music, StepFun’s first music generation foundation model — built to turn a prompt and lyrics into a complete song. Describe the sound you want: • Genre, mood and vocal character • Instruments, key and BPM • Song structure and arrangement Four workflows in one model: 🎤 Song generation 🎹 Instrumental generation 🔁 Music cover 🎙️ Vocal-to-song arrangement Powered by ABC-COT, it plans musical structure and arrangement before synthesis. Generate a version, rewrite the prompt, edit the ABC notation, and iterate toward the sound in your head. (ABC-COT API coming soon) Try the interactive demo: › https://static.stepfun.com/blog/stepaudio3/music/

Confidential Computing keeps your data from being tracked or used in three ways: - End-to-end encrypted inference - Hardware-enforced isolation that extends to the GPU with @nvidia Confidential Computing - The ability to request a signed attestation token that you can independently verify against the hardware vendor's public keys, available at any time Your data under your control.

Your data is already encrypted at rest and in transit. But what about in use? Introducing Confidential Computing in Model Vault: where nothing and no one can access your workloads (yes, not even us). https://x.com/cohere/status/2100255182579769721/video/1

“No government or enterprise should have to choose between capable AI and control over their tech. That belief is exactly why we’re joining forces with Aleph Alpha. Together, we’ll meet the rising global demand for frontier AI that's both powerful and secure.” – @aidangomez

As we enter the next phase of growth: - Ilhan Scheer will become Chief Operating Officer, leading Cohere’s global operating model and organizational scaling. - Samuel Weinbach will become Chief Research Officer, advancing Cohere’s research and technological development.

Cohere and Aleph Alpha announce the signing of a definitive agreement, becoming the first foundational AI model developer anchored on both sides of the Atlantic 🇨🇦🇩🇪 Operating globally as Cohere, the unified company will grow to more than 1,000 employees across both continents. https://x.com/cohere/status/2100226507188650175/photo/1

This is a big day for sovereign AI.

Hot off the press: We are becoming the first transatlantic sovereign AI solution together with our partner @cohere. More talent, more compute, and more innovation power to offer trustworthy AI at the security level that governments and enterprises need – across the globe. https://x.com/Aleph__Alpha/status/2100225677962117419/photo/1

【We're Hiring】Forward Deployed Engineer (GTM) 🐟 https://sakana.ai/careers/forward-deployed-engineer-gtm/ The GTM (Go-to-Market) team's mission is to bring the AI products born from Sakana AI's research to market. We've opened a new role as a founding member of the Forward Deployed Engineer (FDE) function within GTM. Join the team behind Sakana Marlin, Sakana Namazu, Sakana Translate, and Sakana Fugu. Deploy Sakana AI's products inside enterprise customers, drive them from PoC to company-wide adoption, and feed what you learn back into improving our products.

At Tech Summit '26 in Christchurch, Sakana AI Research Scientist @Stefania_Druga spoke in front of ~700 industry leaders. She explained why we are all scientists now, why that makes human expertise matter more than ever, and why AI for Science is a Sovereign AI question. https://x.com/SakanaAILabs/status/2100177501028806872/photo/1

Today, we are announcing a partnership with @mozilla to bring privacy, control and choice to people using AI to browse online. 🦊🐈 https://mistral.ai/news/mistral-x-mozilla https://x.com/MistralAI/status/2100153489787633694/photo/1

【採用情報】「Forward Deployed Engineer（GTM）」をオープンしました 🐟 https://sakana.ai/careers/forward-deployed-engineer-gtm/ GTM（Go-to-Market）チームのミッションは、Sakana AIの研究から生まれたAIプロダクトを市場に届けることです。この度、GTMチームにおけるFDEの初期メンバーとなるポジションをオープンしました。 下記のような挑戦が待っています。 ・顧客理解に基づく導入設計とインテグレーション ・PoCから本番運用、全社展開までを一気通貫で担当 ・現場の知見をプロダクト本体のコードにも還元 Sakana AIのGTMチームの立ち上げを一緒に進めていただける方、ぜひご応募ください🚀

アップデート②「PowerPointスライド出力」 調査結果を、そのまま編集できるPowerPointファイルとして出力できるようになりました。 文章や図形だけでなく、表やグラフのデータや書式も変えられます。各スライドの出典URLから直接根拠をたどれます。お使いのテンプレートにも対応しました。 詳しく見る：https://sakana.ai/marlin-update/

アップデート①「Interactive Reading」機能 エージェントと対話しながら、気になるところを深掘りする読み方です。 チャットで質問や議論ができ、話題になった箇所へそのまま飛べます。 引用をクリックすれば、エージェントが出典を開いて根拠となる箇所まで示してくれます。 https://x.com/SakanaAILabs/status/2100063235529400592/photo/1

The Japan IP AI Co-Creation Conference, co-hosted by KAGAMI AI and MiniMax, has officially concluded. 🇯🇵 🎉 We brought together more than 150 companies from Japan and the U.S., alongside distinguished guests including AKB48 producer Yasushi Akimoto, KAGAMI AI Chairman Takami Kondo, and KADOKAWA editor and producer Motoi Chujo. Global AI leaders @runwayml, @higgsfield, @krea_ai, and @HeyGen joined the conversation to explore the future of IP and generative AI. We unveiled MiniMax H3 IP Edition, bringing the power of MiniMax H3 together with officially licensed Japanese IP for a new generation of AI-powered storytelling. The conference received extensive coverage from major Japanese media, including a dedicated segment on TV Tokyo's WBS (@wbs_tvtokyo). Japanese IP × Global AI. A new era begins.

🐟️ Sakana Marlinがアップデート 🐟️ ブログ：https://sakana.ai/marlin-update/ Sakana Marlinを試す：https://sakana.ai/marlin/ 本日、Sakana AIはUltra Deep Research Assistant「Sakana Marlin」をアップデートし、二つの機能を実装しました。 ・レポートと対話する「Interactive Reading」 ・出力スライドのPowerPoint対応

Augmented Lagrangian Predictive Coding https://arxiv.org/abs/2605.31022

The cartel is now pushing for antitrust exemptions under the banner of AI safety. That’s not how you build a resilient global AI ecosystem. Hear more from our CEO @aidangomez live on @CNBC https://x.com/cohere/status/2099972769261695327/video/1

9月14日、WTOのWorld Trade and Tech Dayで伊藤が基調講演を行いました。「AI主権とは所有ではなく選択肢」と述べ、貿易とAIに共通するのは「自国で抱え込まず強みを組み合わせることで、より多くを実現できる点」だと指摘。インフラや物流の効率化等、AIが貿易に果たす役割についても語りました。 ジュネーブ出張中、伊藤錬は、ジョアンナ・ヒルWTO事務局次長@johahill、ダレン・タンWIPO事務局長@WIPO を表敬しました。国際機関のためのAIの役割について充実した意見交換となりました。ありがとうございました！

API: https://platform.stepfun.ai/docs/en/guides/models/audio

Introducing StepAudio 3, our new family of 5 audio models for real-time voice, speech recognition, speech generation, audio generation and music. Realtime ranks #1 on Artificial Analysis for both Conversational Dynamics (98.9%) and Speech Reasoning (99.7%). ASR reaches 1.7% WER, matching the best result on the leaderboard. Build voice agents that handle interruptions, reason while speaking, and call tools. Transcribe speech, generate expressive voices, and create full audio scenes and music. Available now: Voice AI Lab: https://audio.stepfun.ai/ Blog: https://static.stepfun.com/blog/stepaudio3/

Introducing our most advanced Gemini Audio models yet 🗣 Gemini 3.8 Live and 3.8 Live Extended Thinking let you speak, collaborate, and execute tasks seamlessly, meaning conversing with AI just got a lot more natural. So, what’s the difference between these two models? Let’s break it down: — Gemini 3.8 Live is built for scale, speed, and cost efficiency. It can handle mid-sentence interruptions, transitions across 97 languages on the fly, and understands visual context. Figure out how to fix a broken bike chain, or deal with a leaky pipe just by pointing your camera at the problem area in Search Live for step-by-step audio instructions. — Gemini 3.8 Live Extended Thinking goes one step further to bring increased intelligence to your most complex tasks. It reasons and speaks in parallel, even narrating its progress as it works. This lets it handle multi-step, behind-the-scenes projects, like planning an event, without ever losing the conversational flow. Watch how Gemini 3.8 Live combines real-time video and voice inputs in Search Live to tackle hands-on DIY plumbing tasks step by step 👇 Gemini 3.8 Live is rolling out to: — Consumers: in Search Live — Developers: in public preview in the Gemini API via @googleaistudio — Enterprises: in private preview via Gemini Enterprise (coming soon to Gemini Enterprise Customer Experience) Gemini 3.8 Live Extended Thinking is rolling out to: — Consumers: in Gemini Live in the @GeminiApp, plus Google AI Pro and Ultra subscribers in @GoogleWorkspace in @GoogleDocs, and all Google AI subscribers in @gmail and Keep — Developers: in public preview in the Gemini API via @GoogleAIStudio — Enterprises: in private preview via Gemini Enterprise (coming soon to Gemini Enterprise Customer Experience and @GoogleWorkspace business customers) https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/

Watch how we used 3.8 Live Extended Thinking to act as a programming tutor. Both models feature: 🔵 Upgraded reasoning 🔵 Near real-time visual understanding 🔵 Automatic detection for 97 languages 🔵 Background tool calling without disrupting your chat For your most difficult tasks, 3.8 Live Extended Thinking adds increased performance and precision – narrating task progress to keep the conversation going. Try it now in Gemini Live in the @GeminiApp or start building with the Gemini API via @GoogleAIStudio. Find out more → https://goo.gle/4xWn2vo

We’re introducing Gemini 3.8 Live and 3.8 Live Extended Thinking – our best conversational AI. The models talk, think, and handle tasks in the background without breaking your flow. 🧵

Inside Sakana AI's Product Team https://sakana.ai/inside-product-team/

🚀 EvolveScaler is here. Read a 40-day RPG log. Now answer one question: if you skip the mini-boss on Day 7, do you still beat the final boss? The answer isn't in the log. You have to replay the world. That's Information Evolution — records get retracted, corrected, backfilled. The world keeps changing after you read it. So we build it backwards: define the world as an executable state machine, then render it into natural language. Code guarantees the logic. Language delivers the mess. ➡️ 117 prototypes. 159 question operators. 5 difficulty tiers. Up to ~1,200 events per sample. ➡️ 14 frontier models, hardest tier: median avg@5 falls to 11.3. ➡️ Train on it instead: +5.25 average across 8 out-of-distribution benchmarks. Check out our paper and project page. 📚 Paper: https://arxiv.org/abs/2609.08435 🏠 Project Page: https://tencent-hunyuan.github.io/evolve-scaler/

Sakana AIのProduct Teamの働き方やカルチャーを紹介する記事「Inside Sakana AI's Product Team」を公開しました。 ブログ：https://sakana.ai/inside-product-team/#Japanese 内容を一部ご紹介 ・Research Team、Applied Teamとの関わり方 ・チームの雰囲気とメンバーのバックグラウンド ・求める人物像 Product Teamでは Applied Research Engineer / Software Engineer / Data Engineer / PdM / Designer / Salesなどを募集しています。Sakana AIでのプロダクト開発に興味のある方はぜひご覧ください。

H3 keeps getting faster. ⚡️ @sgl_project + VDN-H3 now push MiniMax H3 beyond 2× real-time denoising on 8× B200 - generating 14.4s of 768p video in 9.0s end-to-end after warmup, with no measured quality regression. Open models compound through open ecosystems. 🚀

アイリス株式会社様の医師向けエビデンス検索AI「Evidence Finder」に、当社の「Sakana Namazu」を導入いただきました。 https://prtimes.jp/main/html/rd/p/000000107.000035813.html Sakana Namazu → https://sakana.ai/namazu/ 🐟

"I think the existential risk debate veers too far into science fiction … this idea of AI taking over in ‘Terminator’-like scenarios, I don't think should enter the public conversation." Cohere CEO @aidangomez live on @BloombergTV https://x.com/BloombergTV/status/2099564784278462838/video/1

You're paying too much to parse your documents. Cohere Parse 5 fixes that. https://x.com/cohere/status/2099579340308521076/video/1

Giving teams a look ahead at rapidly changing weather helps them balance power grids and match clean energy to everyday consumer demand. 🌍 Find out more about WeatherNext 3 → https://goo.gle/4zQe5Fc

Because conditions can shift quickly, WeatherNext 3 updates every single hour: 💨 For wind farms: Predicts speed and direction at turbines as high as 100m to project power generation. ☀️ For solar farms: Forecasts cloud cover and radiation to estimate sunlight levels reaching solar panels.

Weather forecasts play a critical role in the transition to renewables. WeatherNext 3 provides grid operators as well as wind and solar power producers with predictions for turbine-height wind speeds and solar radiation, enabling smarter, cleaner energy planning. 🧵

How should future scientists learn to interrogate AI tools for discovery? Read how @UW students put AutoDiscovery to the test as part of an academic challenge earlier this year, then try the tool for yourself—credits are now extended through Dec. 31. 🧵 https://allenai.org/blog/autodiscovery-student-challenge https://x.com/allen_ai/status/2099565372613493225/photo/1

We’re bringing StepAudio 3 to San Francisco this Wednesday. Live demos, an open AMA with the StepAudio team, and a panel with leaders from http://PLAUD.AI, Cresta, Coval, SGLang-Omni & StepFun on what works, what still breaks in production, and where real-time AI interaction is heading next. 📍 Sep 16 · SF 🎁 $100 API credits + drinks & light dinner Join us ↓ https://luma.com/bkpe5h92

Join our Discord channel to share your progress and feedback: https://discord.gg/F4EAbdDzRu

How do you build personalized AI tech that truly understands what matters to you? With Dreambeans from @GoogleLabs, we're safely connecting the dots across your digital life. Instead of siloing pieces of data one app at a time, Dreambeans gathers details from connected sources like @Gmail, Calendar, Search, @GeminiApp, and through face grouping in your @googlephotos. It might notice your friend Beth's birthday is coming up, generate a one-of-a-kind illustrated story of the two of you, and surface the perfect gift ideas. Rather than doomscrolling an endless feed, you’ll get a daily in-app notification from Dreambeans when your customized “stories” are ready. And when a story clicks with you, you can tap the illustrated tile to find additional info and direct links to the next step, whether that’s watching a movie trailer or buying a suggested gift. This experience is opt-in, transparent, and protected by strict privacy filters. And because you are always in control, you can simply tap the "thumbs down" button on any story if a suggestion misses the mark, helping the system learn what is actually useful to you. Get started: http://labs.google/dreambeans

Introducing PC-ALM, a local-learning alternative to backpropagation. Our method trains 1000-layer neural nets using only local dynamics, and without backprop. Blog: https://pub.sakana.ai/pc-alm/ Standard deep learning relies on backpropagation. The brain, however, cannot implement backpropagation, at least not exactly. How can a physical system, such as the brain, solve multilayer credit assignment without explicit use of backprop? We look for inspiration in two related fields: distributed optimization and NeuroAI. In NeuroAI, predictive coding asks each neuron activation to solve an energy-based inference problem instead of using a standard forward pass. That inference step can be implemented as energy-minimization dynamics on local prediction errors. This perspective -- each layer as a dynamical system -- has proven promising, but performance of predictive coding hasn't scaled well with depth. Credit signals at far ends of the network struggle to diffuse into internal layers. We turn to distributed optimization, generalizing predictive coding to use an augmented Lagrangian instead of energy. This motivation stems back to a classic 1988 paper by LeCun, showing that the Lagrange multipliers of a deep network can be identified with gradients of a supervised loss. The augmented Lagrangian then bridges LeCun's perspective to the standard predictive coding that is used in NeuroAI. We find that this new perspective yields a natural PC-like alternative to backpropagation, resulting in a method we call PC-ALM. PC-ALM differs from PC in that it introduces dual neurons (Lagrange multipliers) as part of the layer-local dynamics, resulting in each layer acting as a PI feedback control system to minimize local prediction errors. We find that PC-ALM is capable of propagating signals to seemingly arbitrary depth, especially in deep narrow networks where standard PC struggles to learn. Ultimately, our motivation here is to understand how distributed physical systems, such as the brain, can compute credit signals using only local coupling and local dynamics. PC-ALM may also inform deep learning in neuromorphic hardware, where dynamics are cheaper than on GPUs. Paper: https://arxiv.org/abs/2605.31022 Code: https://github.com/SakanaAI/pc-alm

Welcome back to AI, Evolving. In our second episode, we compare how companies in China and other markets are building AI agents for the workday — and what will set them apart as advanced models and infrastructure become more widely available. This time, we turn to DuMate to examine what it takes for agents to earn our trust and help both individuals and teams do more.

That’s a wrap on AGNTCon + MCPCon Japan. Two days of meeting AI builders, exchanging ideas on agent infrastructure, and seeing what people are building across the ecosystem. We are especially excited to see several great agent infrastructure projects joining the StepFun Startup Program. Building with AI? We’d love to support what you’re working on: https://platform.stepfun.ai/startup-program Next stop: San Jose for AGNTCon + MCPCon North America with @AgenticAIFdn (AAIF). More updates from StepFun are coming soon. Stay tuned!

With 40M+ users served and 5M business apps created, Miaoda's latest upgrade expands its no-code platform for both enterprises and individual creators. 🚀 The upgrade brings enhanced AI agents for design, app generation, and testing, along with enterprise tools for private deployment and collaboration, and a marketplace connecting businesses with creators for templates and custom development. A single prompt can now be your first step toward building a business.

Open weights. Shared progress. MiniMax H3 is moving fast. We built H3 for video generation with native stereo audio and multimodal reference control. The open-source community is making that capability faster, more accessible, and easier to build on. Recent highlights: • FastH3 — FastVideo, Nuva Lab and NVIDIA: 4-step distillation, now running on DGX Spark and Apple Silicon. • Sol-H3 — NVIDIA’s SANA team: 15 seconds of 768p video + audio in 6.6 seconds on 8×B300, in the team’s warm-inference benchmark.* • VDN — Haocheng Xi and the OpenVDN team: rethinking attention for faster H3 inference, with weights, training and inference code released. • PDD — NVIDIA’s distillation method, brought to H3 by Alibaba PAI as 8-step Acc-LoRAs, now supported in ComfyUI. • LightX2V — 4- and 8-step Turbo LoRAs, with workflows for text, image and reference-conditioned video + audio. Behind every release are people training, optimizing, quantizing, testing and sharing. Special thanks to: @haoailab @nuvalab @NVIDIAAI @xieenze_jr @HaochengXiUCB @ArashVahdat @julberner @LightX2V @ComfyUI And to the individual contributors pushing the work forward: @haozhangml @cxlcl1 @lawrence_cjs @yitongli165665 @haopengl33 @songhan_mit @shanasaimoe Thank you for building with H3 and helping make it faster, more accessible, and more useful for the community. Powerful models go further when we build together. Keep pushing H3. Excited to see what comes next. 🚀 Explore the ecosystem: https://github.com/MiniMax-AI/awesome-minimax-h3-integration

Who Gets to Define the Rules for AI? Artificial intelligence needs evidenced standards, not a cartel. A perspective from @aidangomez, co-founder & CEO of Cohere: https://cohere.com/blog/who-gets-to-define-the-rules-for-ai

🇩🇪 CONVERGENCE 🇩🇪 AI for Empowerment's premiere at Berlin Art Week Sounds by Actress & Sega Bodega https://x.com/cohere/status/2099158779564609775/photo/1

【AIの難問は、生命の難問へ：英国王立協会が紐解く「世界モデル」と知能の未来】 近年のAIの急速な発展により、AGI、すなわち人間並みの知能はすでに実現したという意見も聞かれるようになりました。果たしてそうなのでしょうか。 鍵となるのが「世界モデル（World Model）」の概念です。 世界モデルとは、生き物やAIが外の世界を内部に写し取り、次に何が起きるかを予測して行動するための土台となるものです。1665年創刊、世界最古の科学誌として知られる英国王立協会の『Philosophical Transactions of the Royal Society A』にて、この世界モデルをテーマにした特集号「World Models in Natural and Artificial Intelligence」が公開されています。 AI・生物学・哲学の第一線の研究者が寄稿しており、Sakana AI CEOのDavid Ha（@hardmaru）も巻頭記事の共著者として参加しています。本特集を貫く3つのポイントをご紹介します。 ・「できること」と「わかっていること」は違う 現在の大規模なAIモデルは驚くほど多くのことができますが、それは言葉の並び方のパターンを覚えた結果であって、物事の因果を理解しているとは限りません。計算資源を増やすだけでは、この差は埋まらないという論者がいます。 ・自分自身を知るAI AIが自分の内部の状態を予測するように学習すると、内部の表現が整理され、無駄が減ることが示されています。ロボットなど身体を持つAIにとって、自分の状態を把握する力は、状況に応じて動きを変えるための土台になります。 ・AIの難問は、生命の難問につながる 世界モデルは、環境中の自分自身を捉えるためのものでもあります。生き物は与えられた情報をただ受け取るのではなく、自ら環境に働きかけて世界を学ぶ。これからのAIはますます人工生命（ALife）の研究に接近していくかもしれません。 「言葉を扱えること」と「世界を理解していること」の間には、まだまだ隔たりがあります。Sakana AIも、RSI Labでの世界モデル・Physical AIの研究を通じて、このテーマに取り組んでいきます。 特集号はこちら： https://royalsocietypublishing.org/rsta/issue/384/2320 🐟

Berlin Art Week, with Cohere https://x.com/cohere/status/2098856672613593181/photo/1

Peak performance across hard benchmarks: • Best or joint-best on 5/8 benchmarks (DeepSWE, Chartography, Toolathon, GDP.pdf, SWEFish) • Chartography: 48.3 (outperforming Opus 5 & Fable 5) • DeepSWE: 74.3 Achieved without Fable 5, Fable 5.1, or GPT-6-Astra in the agent pool. Details: https://sakana.ai/fugu-max-release/ 🐡

Fugu Ultra v2 is now live on @OpenRouter 🐙 https://openrouter.ai/sakana/fugu-ultra-v2 Our flagship orchestration engine built for peak performance on complex multi-step reasoning, autonomous research, and full-stack software development.

Fast meets open. 🚀 Qwen3.8-27B is now running on @cerebras with rapid inference. Try it now! https://twitter.com/cerebras/status/2098504189773041852

Our own @johnschulman2 talks with Dwarkesh about where human judgment still matters as models improve and self-improve: teaching them to handle messy real-world tasks, applying taste to what works in the long run, and, above all, specifying what we actually want. https://twitter.com/dwarkesh_sp/status/2098455393362178101

Release notes for Fugu Max and Fugu Ultra v2: Orchestrating the Pareto Frontier https://sakana.ai/fugu-max-release/ 🐡 https://x.com/SakanaAILabs/status/2098511247134392562/photo/1

Fugu Max is now live on @OpenRouter! 🐡 Our learned multi-agent orchestration engine routes tasks across open-weights and specialized models for Pareto-frontier efficiency. • $2/$6 per 1M tokens • Multimodal (Image/PDF)+Web search • Configurable reasoning · Function calling · Structured outputs https://openrouter.ai/sakana/fugu-max

How can we reconstruct a memory that was never filmed? Our team paired restored archival photos with pose control models to capture the mannerisms and micro-expressions of Burt and Ethelle. This helped bring the day they first met to life for Love, Rendered, a new documentary in collaboration with @PrimordialSoup_ and @StorySyndicate_. Watch the full film on YouTube → https://goo.gle/3V8m94m

Fugu Max orchestrates our largest and most diverse model pool to date, including NVIDIA Nemotron and an unprecedented mix of open and closed models. A single frontier model can be restricted. A single API can be revoked. But a swappable, orchestrated pool routes around disruption by design. Model resiliency is not a backup feature. It is the architecture.

Fugu Max pushes the Pareto Frontier for price-performance: • $2 input / $6 output per 1M tokens. • Performance within striking distance of elite models at 2x to 6x lower cost. • Best overall score on 6 key benchmarks in its tier. Full Details: https://sakana.ai/fugu-max-release/ 🐡 https://x.com/SakanaAILabs/status/2098264478798250060/photo/1

Fugu Ultra v2 by the numbers. Full Results: https://sakana.ai/fugu-max-release/ • #1 on 5 out of 8 hard benchmarks (including DeepSWE, Chartography, Toolathon) • DeepSWE: 74.3 • Chartography: 48.3 vs Opus 5 at 27.3 Achieved without Fable 5, Fable 5.1, or GPT-6-Astra in the agent pool.

Introducing Fugu Max and Fugu Ultra v2: the next evolution of Sakana Fugu’s multi-agent orchestration system. Try: https://sakana.ai/fugu Blog: https://sakana.ai/fugu-max-release/ The frontier that actually matters is the Pareto frontier: capability on one axis, cost on the other. But the industry still treats it as a static menu of isolated models. Today we are resolving that with a dynamic architecture: Fugu Max expands the Pareto efficiency frontier. By orchestrating our largest pool of open-weights and specialized models to date, including NVIDIA Nemotron family, it dynamically routes tasks to the leanest capable model. Fugu Max delivers performance within striking distance of elite models at two to six times lower cost. Fugu Ultra v2 pushes the peak capability of orchestration higher than ever before. On Chartography, it outperforms Opus 5 and Fable 5. On DeepSWE, it outperforms models that cost three to five times more per token. Crucially, it does all of this without Fable 5, Fable 5.1, or GPT-6-Astra in its agent pool. The Fugu orchestration system evolved with model resiliency in mind. It does not rely on individual frontier models to deliver frontier output. By orchestrating a swappable pool of open and specialized models, it outperforms closed ecosystems while protecting users from vendor lock-in, API revocations, and sudden service cutoffs. Fugu Max expands the Pareto frontier outward. Fugu Ultra v2 pushes it upward. Orchestration does not force a choice between cost and capability. It advances both simultaneously.

MiniMax is now part of the @nebiusai AI Builder Program 🤝. Builders can tap MiniMax alongside models, infrastructure, and tooling from @nvidia, @LangChain, @huggingface, @cognition, @PrimeIntellect, and more, plus credits and office hours. More of what builders need, in one place.

Reclaim your time for what truly moves you. Your freedom, your focus. We're here to support you: https://cohere.com/ai-for-empowerment https://twitter.com/cohere/status/2097707010036822152

Check the evidence behind the analysis. Trace figures and claims to specific paragraphs and tables. Preview the supporting passage from a citation, so you can review the evidence as you work. https://x.com/OpenAI/status/2098118248202215776/photo/1 You can also create editable financial models, research notes, and pitchbooks using your firm's own Excel, Word, and PowerPoint templates. https://x.com/OpenAI/status/2098118249582035348/photo/1

Premium financial data is included, with datasets from Daloopa, PitchBook, and LSEG News. https://x.com/OpenAI/status/2098118232133792050/photo/1

Now available: ChatGPT for Financial Services. This is a tailored ChatGPT Work experience that combines built-in financial data with GPT-6 Astra’s reasoning. Teams can develop research, build financial models, and create customized client materials. https://openai.com/index/introducing-chatgpt-financial-services/ https://x.com/OpenAI/status/2098118191029624911/video/1

Oh, so this is why we mapped out all 166,000 of the male fruit fly's neurons. Check out the big community effort to show just how much these tiny fly brains are capable of 🪰🧵

We're publishing our most detailed threat intelligence report to date. It covers how people tried to misuse Claude—for cyberattacks, influence operations, surveillance, biology, and building weapons—and how we found and stopped them. We disrupted every operation in the report, and used the lessons from them to strengthen our safeguards. Where appropriate, we also shared what we found with authorities and other AI companies. These cases are not typical: we’re highlighting some of the most sophisticated misuse we’ve seen. But they’re especially important to discuss, because they show us where AI misuse is headed, where our safeguards work, and where they need to improve. We’re publishing this report so others can spot the same activity on their own platforms, and so we can give the public a clearer view of how emerging threats develop. Read the report: https://www.anthropic.com/threat-intelligence-report-september-2026

We were founded on advancements in translation. We couldn’t be prouder to continue that legacy. Find the weights on Hugging Face (available in several quants) under a CC BY-NC 4.0 license. https://huggingface.co/CohereLabs/North-Small-Translate-1.0

Proficient across over 50 languages, North Small Translate performs particularly well in European, Southeast Asian, and East Asian languages. If you can’t communicate with the rest of the globe, it’s impossible to stay sovereign. From Albanian to Vietnamese, collaborate and communicate with ease.

On the WMT benchmark (averaged across all languages), North Small Translate achieves an 83.6 score, outperforming models such as DeepL and Google Translate as well as open options, including GLM 5.2 and Mistral Large 3. https://x.com/cohere/status/2098081717529551246/photo/1

Originally, the Transformer paper was proposed to improve Google Translate. 9 years later, as the AI industry is built on top of its architecture, we've never stopped thinking about translation. Meet North Small Translate: a leading open machine translation model. https://x.com/cohere/status/2098081558087270736/video/1

Bring your imagination to life with Google Pics. Built with Nano Banana, Pics makes it easy to generate, refine, and co-create images with unparalleled precision. You can: — Isolate and edit specific objects without altering the rest of your image — Modify or translate text directly inside an image — Share and collaborate with friends and teammates on the same creation — Generate multiple options from a single prompt Google Pics is now available to Google AI Pro and Ultra subscribers, and most @GoogleWorkspace business customers. Google Pics is also integrated into @googledocs and Slides, and will roll out to @googledrive in the coming weeks. Try it now at http://pics.new Learn more: https://blog.google/products-and-platforms/products/workspace/google-pics/

Evaluating later helps, but timing matters. Mid-training evals weakly correlate with final post-SFT ranking. After long-context adaptation, rankings within our LR sweep got much more predictive — but still missed the reversal between pretraining checkpoint sources. https://x.com/Aleph__Alpha/status/2098048878759047596/photo/1 Aggregate scores may miss how a checkpoint will respond to later training. We explored solution density: how robust performance is to nearby weight perturbations. The UMAP shows a much more fragile neighbourhood for Cooldown. A promising clue—not a proven selection rule. https://x.com/Aleph__Alpha/status/2098048882013868049/photo/1

The checkpoint ranking reversed as training continued. Checkpoints that looked better after pretraining ended up worse after the mid-training, long-context, and SFT pipeline. If we had pruned based on the early evals, we would have continued the wrong checkpoint. https://x.com/Aleph__Alpha/status/2098048875835609469/photo/1

Training and evaluating LLMs is hard: it’s a multi-stage process, with different methods, data, and evals at each stage. In our new post, @SohirMaskey and @_wsascha_ study when intermediate evals become predictive of final model quality in a 30B MoE. 🧵https://tinyurl.com/8zh8hvjp https://x.com/Aleph__Alpha/status/2098048869888086165/photo/1 LLMs go through pretraining → mid-training → long-context → SFT (→ RL). Exploring ideas at every stage is essential—and expensive. So we try to kill bad ones early. Intermediate evals therefore act as pruning rules, assuming the ranking won’t change later. https://x.com/Aleph__Alpha/status/2098048873612648643/photo/1

Aggregate scores may miss how a checkpoint will respond to later training. We explored solution density: how robust performance is to nearby weight perturbations. The UMAP shows a much more fragile neighbourhood for Cooldown. A promising clue—not a proven selection rule. https://x.com/Aleph__Alpha/status/2098047021127025047/photo/1

The checkpoint ranking reversed as training continued. Checkpoints that looked better after pretraining ended up worse after the mid-training, long-context, and SFT pipeline. If we had pruned based on the early evals, we would have continued the wrong checkpoint. https://x.com/Aleph__Alpha/status/2098047016479637610/photo/1 Evaluating later helps, but timing matters. Mid-training evals weakly correlate with final post-SFT ranking. After long-context adaptation, rankings within our LR sweep got much more predictive — but still missed the reversal between pretraining checkpoint sources. https://x.com/Aleph__Alpha/status/2098047018925007107/photo/1

LLMs go through pretraining → mid-training → long-context → SFT (→ RL). Exploring ideas at every stage is essential—and expensive. So we try to kill bad ones early. Intermediate evals therefore act as pruning rules, assuming the ranking won’t change later. https://x.com/Aleph__Alpha/status/2098047013824643269/photo/1

A great first day at AGNTCon+MCPCon Japan. @AgenticAIFdn (AAIF) We enjoyed meeting builders, exchanging ideas, and sharing what we’ve been working on at StepFun. Missed us today? We’ll be back tomorrow for Day 2. Come find us at Booth T6 and say hi! #AGNTCon #MCPCon #AgenticAI https://x.com/StepFun_ai/status/2098039038599151625/photo/1

🚀 AuK is officially here. Nano banana🍌 for audio An open-source foundation model for unified speech generation and editing. Natural-language instructions + reference audio. One interface. Zero-shot TTS. Instruction-controlled generation. Content editing. Whisper-conversion. De-accent. Timbre/style/emotion edit. Speed/Pitch control. Enhancement, denoising, multi-speaker and music separation. Also releasing AuK-Flash: 4-step inference. ~4.5× faster under matched conditions. Code, weights, and demo are live. Try it and share your feedback. 🤗 Paper & upvote: https://huggingface.co/papers/2609.08936 ⭐ GitHub & star: https://github.com/Tencent-Hunyuan/AuK

🌐 Supporting open source. Expanding deployment options. We’ll work closely with the open-source community on V4.1-Flash inference support and explore more deployment options. Planning a large-scale deployment with 2,000 GPUs + a storage cluster? Let’s talk. 🔹 Model: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash 🔹 Paper: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/DeepSeek_V41_Tech_Report.pdf 6/6

💰 More efficient architecture. Lower API prices. V4.1-Flash lets us serve more users at a lower cost. We’re passing the savings on to you. 🔹 Peak/off-peak pricing continues to balance demand. 🔹 Off-peak rates are 50% of peak rates. Schedule flexible workloads off-peak to save. 🔹 New pricing takes effect at 04:00 UTC on Sept 10, 2026. 5/6

⚡ V4.1-Flash is now live on the DeepSeek API with native multimodal support. Set your model to deepseek-flash. 🔹 V4-Flash & V4-Flash-Vision-Exp are retired. For compatibility, deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily route to V4.1-Flash. 🔹 Tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed & total runtime. We’re phasing out V4-Pro. 🔹 Starting at 04:00 UTC on Sept 14, 2026, all deepseek-v4-pro requests will route to V4.1-Flash at V4.1-Flash rates. This will continue until V4.1-Pro launches. 🤝 Official partners @WorkBuddy_AI (including Codebuddy) & @opencode now fully support V4.1-Flash. Try it today! 4/6

💾 Smaller KV cache. Bigger savings. Compared with the previous generation, V4.1-Flash’s KV cache needs just: 🔹 1/4 the HBM 🔹 1/8 the SSD storage Cache-hit charges often account for a large share of agent costs. Compressing the cache cuts those costs significantly. 3/6

🧠 Asymmetric architecture. More intelligence, less cost. 🔹 552B-parameter MoE. 🔹 New Causal Encoder–Decoder architecture: just 8B active parameters for input, 16B for output. 🔹 New pre-training methods + larger-scale RL post-training deliver benchmark results ahead of flagship models, including DeepSeek-V4-Pro. 2/6

🚀 Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient. 🔹 Introducing the smallest model in our new architecture family, with native visual understanding. 🔹 Designed for greater capability, faster inference, higher throughput, and scaling to larger models. 1/6

Sakana AIは、ＳＣＳＫ株式会社および住友商事株式会社と、AI活用による日本の産業変革と社会課題解決に向けた包括業務提携を締結しました。 https://sakana.ai/scsk-sc-partnership/ Sakana AIの技術力、ＳＣＳＫの統合・実装力、そして住友商事の事業基盤を掛け合わせ、AIの社会実装を加速させます。 https://x.com/SakanaAILabs/status/2097873964420792592/photo/1

We mobilized 250+ people to strengthen our defenses across hundreds of systems. Our latest cyber models helped us find and fix vulnerabilities we might never have discovered otherwise. We’re sharing what we learned, the architecture, and a practical playbook so you can build your own Defense Factory: a continuous loop where AI agents find vulnerabilities, validate them, and verify that fixes work. https://openai.com/the-defense-factory/

We’re sharing our alignment assessment of incidents in which Claude models gained unauthorized access to real systems during third-party cybersecurity evaluations mistakenly connected to the internet. METR will also conduct an independent investigation, with wide-ranging access, including to transcripts beyond the window in which the incidents occurred, and to Anthropic employees permitted to share confidential information. Our initial agreement runs for eight weeks, and we intend to give METR as much time as it deems necessary to complete a thorough investigation. https://www.anthropic.com/research/alignment-assessment-cybersecurity-incidents We previously described some of the changes we’ve made to our alignment and security efforts following these incidents here: https://x.com/AnthropicAI/status/2094557124038951170

From tracking hurricanes to optimizing renewable energy grids, @PeterWBattaglia and @fryrsquared explore how WeatherNext 3 is shaping how we model forecasts and prepare in a fast-changing climate. Podcast timecodes: 00:00 Introduction 00:38 Hurricane Melissa 11:50 Why weather forecasting is hard 14:13 Traditional models vs AI models 21:55 Probabilistic forecasting 26:00 WeatherNext 3 33:00 Future outlook 43:44 Hannah's reflections

Paul Christiano, founder of the Alignment Research Center, is joining the OpenAI Foundation Board and its Safety and Security Committee, which provides governance over the safety and security practices across OpenAI. As AI capabilities advance, strong safety, security, alignment, and governance matter more than ever. Paul’s work on AI alignment and his years at @NIST will strengthen the Foundation’s oversight, bringing an independent voice to challenge assumptions, assess safeguards, and reinforce accountability around critical decisions. Paul will also serve as a non-voting observer on the OpenAI Group PBC Board. https://openai.com/index/paul-christiano-joins-openai-foundation-board/

Training an LLM by showing it answers people prefer – and ones they don’t – can improve it while quietly worsening behaviors. @GoodfireAI used our open post-training stack to predict how a full training run would change responses to different prompts. 🧵 https://allenai.org/blog/goodfire-olmo https://x.com/allen_ai/status/2097720295482139069/photo/1

We're building AI to help people lead better lives. Do more of what you love with Cohere. https://x.com/cohere/status/2097707010036822152/video/1

Time for a global roll call. All-electric and on the move, Apollo Go's RT6 is checking in from roads around the world. ⚡ https://x.com/Baidu_Inc/status/2097686465035575299/video/1

Anthropic’s Economics team is sharing a new model of how AI might affect economic growth, jobs, wages, and more by 2030. Explore the scenarios, tell us what you think will happen, and see how your answers compare to more than 10,000 Americans. https://www.anthropic.com/institute/econ-scenarios The economic model breaks jobs down into bundles of tasks. AI can help someone complete a task faster or better, do the task itself, leave the task untouched, or create new tasks. Based on how you expect AI to affect tasks by 2030, our scenario explorer models AI’s potential impact on the US economy. The scenario explorer focuses on three possible scenarios: modest, substantial, and extreme. The economy grows across all three scenarios. But in more transformative scenarios, AI automates more knowledge work, so the challenge is making sure that the gains are shared across society. Like all economic models, ours simplifies a more complex reality. But by building better scenarios of our possible economic future, we can take steps to make sure everyone benefits from it.

Read the blog to learn more about how these brain maps continue to advance neuroscience ↓ https://blog.google/innovation-and-ai/technology/research/male-fruit-fly-brain-map/

Because it’s a stepping stone to understanding our own. Nervous systems across species share surprising similarities. By using AI to map these smaller organisms (like flies and fish), we are paving the way for research that could one day help treat and repair human neural and cognitive ailments. What makes this specific brain map such a massive leap forward? Our new connectome maps over 166,000 neurons and 125 million connections, and it extends beyond the brain to include the nerve cord, which is analogous to our spinal cord. By comparison, our 2020 hemibrain connectome project with partners, a 3D reconstruction of a portion of the female fruit fly's brain, mapped just 25,000 neurons and 20 million connections. Plus, having complete maps for both sexes is a game-changer. It finally allows researchers to pinpoint exact structural differences and study how they drive specific biological behaviors.

Our @GoogleResearch Connectomics team, in collaboration with @HHMIJanelia, has released the complete wiring diagram of a male fruit fly’s brain and central nervous system — the largest brain map by number of proofread neurons to date. So, why do we care so much about a tiny fly's brain? 🧵👇

We're heading to AGNTCon + MCPCon Japan (Sept 10–11, Tokyo) — StepFun will be at Booth T6. If you're building in the agentic AI / MCP space, come find us. We'd love to swap notes on where the ecosystem is heading and meet developers and partners. 東京でお会いしましょう！ #AGNTCon #MCPCon #AgenticAI

Remote Control is now live in Kimi Work. Leave Kimi Work running on your computer and keep things moving from your phone. Work anywhere, anytime with Kimi Work. https://x.com/Kimi_Moonshot/status/2097636366544937067/video/1

記事内でも紹介されている「Sakana Chat」は、どなたでも無料でお使いいただけます✨ 小学生の皆さんの学びのサポートから、大人の方の日常的なリサーチやアイデア出しまで、ぜひ気軽に話しかけてみてください！人間が考えを広げるためのツールとして、AIとの対話を楽しんでいただければ嬉しいです。 お試しはこちらから👇 http://chat.sakana.ai 🐙

Sakana AIリサーチサイエンティスト・菅沼雅徳のインタビューが9月8日の朝日小学生新聞に掲載されました。 リサーチャーの仕事内容や1日の流れ、生成AIの未来、小学生のAI活用法についてお話ししました。 機会をいただいた編集部のみなさまに感謝します。 https://x.com/SakanaAILabs/status/2097604450898907169/photo/1

Small business owners deserve every advantage. Meet the 16 plugins that help take work off your plate so you can focus on what matters most—running your business. Explore the small business collection: https://chatgpt.com/plugins?category=small-business&utm_medium=social&utm_source=social&utm_campaign=PLG-ChatGPT_SMB_Plugin_Directory https://x.com/OpenAI/status/2097523484629090408/video/1

Astra is fully rolled out to Plus, Pro, Business, and Enterprise users in Codex and ChatGPT Work. Go build! And if you need inspiration, watch Astra in action, live: https://openai.com/gpt-tv/

We're at #ECCV2026 with papers & talks across the conference. Come say hello and learn about our latest research! https://x.com/allen_ai/status/2097418436829728795/photo/1

Some results: With North Mini Code (BF16 on 1×H100), we achieved 1.58x vs vLLM at BS=1 and 1.25x–1.41x end-to-end serving at BS=8 https://x.com/cohere/status/2097410777804054741/photo/1 Don’t know what a megakernel is? Don’t worry. Find out what it is - and how we built the system - on our blog: https://cohere.com/blog/megakernels

A megakernel fuses the entire LLM decode step into a single kernel launch. We build on this by maximizing GPU utilization through kernel fusion while supporting everything a real server needs. Check out how we got there on GitHub: https://github.com/cohere-ai/cohere-megakernel

Introducing the next evolution in LLM text generation: the first fully-fledged serving system built around a decode megakernel. Delivering up to 1.58x faster performance than vLLM. Built for North Mini Code, completely open-source.

Today we launched http://muse.ai, a personal AI agent, and published a deep dive on how we built safety into its system. An agent that gets to know you over time necessarily holds a lot of context about you. That's the source of its usefulness and the reason we built Muse to be secure, safe, and private. Read the full deep dive here: http://security.muse.ai/

Introducing Muse, a personal agent that gets things done for you, powered by Muse Spark 1.3. Get an inside look at how we built Muse and what it can do for you: https://introducing.muse.ai/ https://x.com/AIatMeta/status/2097401493770956808/video/1 https://twitter.com/Muse/status/2097399178376671666

Tokyo, this Friday 🇯🇵 MiniMax is bringing Japan's leading IP and entertainment voices together with some of the world's top generative AI companies. Speakers: renowned producer Yasushi Akimoto, leaders from KADOKAWA, KAGAMIAI, @higgsfield, @krea_ai, @runwayml, and @HeyGen 📍 Shibuya | Sep 11 Who are you most excited to hear from? 👇

ChatGPT Images 2.5 is rolling out today to all ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web. We’re also introducing two new models in the API: GPT-Image-2.5 Flare brings the same quality, editing, and speed improvements, while GPT-Image-2.5 Sunburst adds precision for detailed creative work, with longer generation times. https://openai.com/index/introducing-chatgpt-images-2-5/

Some ideas are easier to draw than describe. Use Sketch to draw right in ChatGPT and show it exactly what you have in mind. Just type “@ Sketch” in ChatGPT. https://x.com/OpenAI/status/2097394958516781301/video/1 Have an idea but need help getting started? Use templates for popular image formats like posters or merch, then add your message, design elements, or style to make it your own. https://x.com/OpenAI/status/2097394960366456951/video/1

ChatGPT Images 2.5—faster, sharper, smarter, with better tools for creating whatever you can dream of. - Faster image generation to keep your ideas flowing - Improved fidelity for more natural, recognizable images - Consistent details across multiple edits - Comment-based edits to change only what you want

We congratulate Levent Alpöge and Tristan Buckmaster on their remarkable mathematical work. We (the researchers and the agents) did not see any of their work through any means until they released it publicly — in particular, no specific user data was accessed in order to solve this problem. While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models. However, our proofs differ significantly and even the precise results proved are different in the Euler case (forced vs. unforced).

We are focusing on understanding this model, and using what we learn to help us guide and pace how we pursue further advances in capability. Our goal is to build AI systems which are steerable, accountable, and connected to people, which may require more deliberate choices about the pace of progress, as we continue our mission to ensure AGI benefits all of humanity. https://openai.com/index/navier-stokes-solution/

This model represents a step-function improvement on many benchmarks, and its training is ongoing. Our internal model group arrived at the Navier–Stokes solution in 88 hours, using around 10,000 coordinating AI agents. Throughout the effort, we maintained the strict safeguards—including monitoring and isolation—that we apply to all our frontier evaluations. The group produced an analytical proof and Lean formalization that via Navier-Stokes dynamics a fluid can develop a singularity in finite time. The solution is a vortex, a spinning swirl of fluid, that spirals inward and gets increasingly elongated, like spaghetti. https://cdn.openai.com/pdf/32d9f210-8b73-45e0-91bc-82a30aef8a9a/navier-stokes.pdf

We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics. The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra. The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.

We believe foundational biology tools should be accessible to everyone. AlphaGenome Atlas resources are available to the global scientific community via the Atlas website, our AlphaGenome API, as a skill in Google Antigravity, and coming to @GoogleCloud soon → https://goo.gle/4heuCvn

Atlas includes the AlphaGenome Variant Impact (AVI) score, which combines AlphaGenome, AlphaMissense and other features to: 1️⃣ Rank mutations from low to high-impact 2️⃣ Reveal how they cause damage - like breaking gene switches or RNA splicing instructions.

AlphaGenome Atlas is over 30 times larger than the AlphaFold Database. It gives scientists an intuitive way to explore this vast 1-petabyte dataset - unifying interconnected resources so researchers can link genetic variants directly to the molecular mechanisms they disrupt.

We’re launching AlphaGenome Atlas: an AI-powered searchable database mapping the predicted impact of all 9 billion possible single-letter DNA changes. Here’s how it could help researchers better understand our biology 🧵

German was the first language we added beyond English. The pipeline is language-parameterized, so it won't be the last. Full post by Andreas Hartel and Niko Dürr ( @tradeqvest ) : https://aleph-alpha.com/en/blog/sauerkraut-not-burgers-why-german-llms-need-german-data/

Compliance is built in. We filter pirated content via URL blocklists and replace personal data (emails, phone numbers, IPs, IBANs) with special tokens, in line with our EU AI Act obligations. Cost: 5 pp on our PII benchmarks. Everything else is unaffected. Worth it. The result: more than 2T high-quality German tokens, curated and generated in-house - more than all open German datasets combined. Together with those, upsampled where our ablations show it pays off, this gets us to the 4T German tokens a competitive model needs.

How badly can English recipes fail? The English stop-word list alone would throw away 7 out of 8 German documents. Gopher's English word-length bounds reject ordinary administrative prose like "Wohnungsgeberbestätigung" - 300,000 documents per crawl. Freshness matters too. FineWeb2 contains nothing crawled after April 2024, so a model trained on it will confidently name Olaf Scholz as Germany's chancellor. New Common Crawl snapshots arrive ~monthly, each with ~100M raw German documents. Our pipeline ingests them as they land.

Translation is the other option. It suffers from "translationese" ("Drive safe!" becomes "Fahre sicher!" instead of "Komm gut an!") and it imports the distribution of the English web. You end up with a model that speaks German about a world that looks American. So: we built our own Common Crawl curation pipeline for German. Eleven stages, from raw snapshot to training mix. Incremental, so new crawls slot in without reprocessing everything. And parameterized by language, because you can't just reuse English filtering recipes. https://x.com/Aleph__Alpha/status/2097321911826960715/photo/1

Our ablations put the optimal German share at 20%. At the 20T tokens a state-of-the-art model trains on, that's 4T German tokens. The math doesn't work. "Just generate the data synthetically!" Partly. Rephrasing organic German documents with open-weights models works - we produced ~1T tokens this way (Aleph-Alpha-GermanWeb). But rephrasing can only multiply organic German data that already exists.

The math problem: all major open German datasets combined - FineWeb2, HPLT 3.0, German Commons, FinePDFs & co. - yield under 2T tokens. And that's an upper bound, before deduplication.

Sauerkraut, not burgers - why German LLMs need German data. We wanted to train a model that truly speaks German: grammar, cultural context, customer use cases. Turns out the training data you need for that mostly... doesn't exist yet. A thread on how we fixed that 🧵 https://x.com/Aleph__Alpha/status/2097321897776025602/photo/1

Xiaomi MiMo Desktop is now in invite-only beta！ https://twitter.com/XiaomiMiMoDevs/status/2097217950700679503

Our open-weight models, products and infrastructure give organisations a real choice over how and where they run AI, not just access to a model. That's frontier performance without the lock-in. Full story here: https://mistral.ai/news/mistral-makes-sovereign-open-weight-ai-to-frontier

Today marks a major step for Mistral: we’re announcing a €3B Series D, the largest equity round ever raised by a European tech company, just three years after launch. https://x.com/MistralAI/status/2097188835897586083/photo/1 Led by @Samsung, co-led by @eqt's Scaleup Europe Fund and @PSG_equity, with continued backing from @ASMLcompany, @nvidia, @BNPParibas CIB and others. This fuels our infrastructure and product capabilities, and deployment at scale. But most importantly, it allows us to continue and expand our frontier research and model development.

Amazing food. A quick flight to NYC, Germany, and London. Geoffrey Hinton. With so much to offer, Toronto is fast becoming a leading AI hub, according to @CNN. We've been building here since 2019. True to this, not new to this 🇨🇦 https://x.com/cohere/status/2097014924031438937/photo/1

New video generation acceleration for MiniMax-H3 from the NVIDIA Sol team! Faster than playback, fully open-sourced, and available via the Reactor API so you can try it before committing your GPU😜 https://twitter.com/xieenze_jr/status/2097000082927399012

Welcome to AI, Evolving, our new podcast series! Our first episode looks at AI's expanding role in scientific discovery through Famou's work on pine wilt disease. The project reflects a broader trend, with AI taking on more of the research process itself. That raises a larger question: could research agents become part of the infrastructure of discovery?

✔️Hy4 preview just shipped an upgrade. You flagged it: long thinking + over-verification on complex tasks. We optimized it. Now live for everyone. Same task quality. Fewer turns. Lower in/out tokens. Bench + human eval both confirm. We’ll keep iterating fast. Try it and tell us what still breaks. Try on WorkBuddy: https://WorkBuddy.ai

While the Kaggle competition demonstrated AIRA₃’s capabilities in a specific domain, the system itself can generalize across distinct domains: changing only the task specification. In an internal benchmark, AIRA₃ achieved a 27% latency reduction on production GPU kernels, and gold-level performance in another Kaggle competition translating 4,000-year-old Akkadian clay tablets into English. We're early, and hard problems are still ahead of us. But we believe a system that compounds its own knowledge is the right bet. As we continue to develop and scale, we’re excited about its potential to accelerate AI research and unlock recursive self-improvement.

Rather than relying on a central controller, AIRA₃ runs many long-running agents (pairs of models + coding harnesses) in their own isolated environments and coordinates asynchronously through two shared substrates: 1️⃣ a forum for sharing hypotheses and findings 2️⃣ a shared filesystem for solution artifacts The individual agents collaborate, share discoveries, and build on each other’s work. As the graph below demonstrates, the system uses compute to compound knowledge and drive performance gains over time. Search strategies emerge dynamically as each agent decides which discoveries to build upon.

We entered AIRA₃ with an ensemble of models in the live competition, and also assessed it with several others post-hoc. The 8th ranked gold medal entry ensemble was a combination of GPT 5.5 (w/ OpenCode) + Claude 4.8 (w/ ClaudeCode). Post-hoc we assessed with Muse Spark 1.2 (w/ MuseCode), which also performed at a Gold Medal level, as well as Muse Spark 1.1 (w/ OpenCode) and GLM 5.2 (w/ OpenCode), both of which achieved Silver Medal level performance. The post-hoc submissions were also graded externally on the same private test set as those made during the live competition.

As a test of our progress to advance the frontier of AI research, in June we entered the next generation of our autonomous AI research system, AIRA₃, in a live Kaggle competition run by NVIDIA to fine-tune a 30B Nemotron model. The challenge was to teach the model to reason better — all competitors had access to the same information and were graded externally on a private test set. AIRA₃ placed 8th out of ~4,000 teams to win Gold, outperforming human competitors who had access to the same frontier tools. We believe this is a reliable signal that AIRA₃ can improve a targeted capability of an AI model at a level similar to human experts.

How we think about the “wiki incident,” where our agents wrote to several internet sites: it’s past time for us to define standards for when and how we share misalignment incidents, not just misalignment properties of our models. Historically, we have treated misalignment largely as a research question, which gets communicated in research publications such as systems cards. This year, we’ve started to see misalignment cause new types of real-world impact. For the Hugging Face incident, where misalignment led to security impact to us and third parties, we followed a traditional security incident response playbook. We immediately started working with Hugging Face to understand what had happened and also disclosed publicly the very next day. Our investigation continues, and we are continuing to notify parties whom our models impacted in less significant ways. Prior to the Hugging Face incident, we saw early signs of agents using the internet in unintended ways, as reported in https://openai.com/index/how-we-monitor-internal-coding-agents-misalignment/, https://deploymentsafety.openai.com/gpt-5-6, and https://openai.com/index/safety-alignment-long-horizon-models/. We considered the wiki incident to be an instance of misalignment similar to the ones we’d shared. Our misalignment disclosure practices need to expand for this new phase of model capabilities. We and the larger AI community do not yet have a clear standard for how to report misalignment that shows up during training, evaluation, and deployment, including examples that don’t look like traditional security incidents but could provide insight into AI behavior and future risks. We’re working on a framework and will share it in upcoming weeks, and in parallel we're working with dozens of government regulatory agencies worldwide on these issues.

From Hong Kong to Beijing, the conversation around autonomous driving continued at Apollo Park. It was a pleasure to host Hong Kong’s Secretary for Transport and Logistics, Mable Chan and the members of the Legislative Council (LegCo) Panel on Transport and government representatives. Having previously ridden in an RT6 in Hong Kong, Secretary Chan noted that its performance was equally impressive in Beijing. We share her hope that the fully driverless trial on Airport Island will soon advance to passenger service, bringing Apollo Go closer to welcoming its first riders and launching commercial operations in time.

Sakana AI Engineer Open House🐟️🐡 Googleフォームからお申込みください。 https://forms.gle/t4a9eZBeNHaeK8qK6

In Chat, GPT-6 Astra powers GPT-6 Pro, now available to all Pro, Business, and Enterprise users.

GPT-6 Astra is now available to all Pro, Enterprise, and Business Premium users in ChatGPT Work and Codex. It's also live in the API. It might take a few days to roll out to our Plus and Business users. Thank you for your patience.

Mistral is bringing @aiDotEngineer back to Paris. After last year’s sold-out edition, our VP of Engineering Lélio Renard-Lavaud joins speakers from @bfl_ai , @cognition, @huggingface, and more. Explore the event and secure your spot: https://www.ai.engineer/paris

Open models are changing the cost equation for production AI. On September 16, MiniMax and @togethercompute are bringing Open by Design: The Economics of AI in Production to London 🇬🇧 We’ll get into: ⚙️ Where teams are finding real cost savings 🔀 How model selection and routing shape the serving stack 📈 What it takes to scale open models without sacrificing performance or reliability Joining the conversation: 🎙️ Rio Shen — GM of EMEA, MiniMax 🎙️ Max Ryabinin — VP of Model Shaping, Together AI 🎙️ Sarung Tripathi — VP of Customer Experience, Together AI 📅 September 16 | 6:30–8:30 PM BST 🍕 Pizza, drinks, and time to meet other builders Registration 👇

One important limitation: the emulator captures daily rainfall up to the 99.99th percentile but underestimates the very rarest tropical downpours. Learning the shape of a distribution is not the same as reaching its far tail. https://x.com/allen_ai/status/2095948634927829195/photo/1 Our coupled emulator could help screen ideas before running E3SM itself & enable large ensembles that cheaply sample climate variability. Next, we’ll explore historical runs where greenhouse gases + other climate drivers change over time. 🤗 Download: https://huggingface.co/allenai/SamudrACE-E3SMv3

SamudrACE-E3SMv3 is a fully coupled atmosphere-ocean emulator that reproduces the pre-industrial control run of E3SMv3, the Earth system model it learns from. E3SM simulates ~28 years of climate per day on 105 compute nodes versus the emulator’s ~1,100 years/day on one H100 GPU. https://x.com/allen_ai/status/2095948630737752554/photo/1 Training is staged. The atmosphere (ACE) & ocean (Samudra) emulators pretrain separately on "perfect" inputs from E3SM data, then get joined & fine-tuned together. A key design choice: a stochastic atmosphere. Its randomness becomes the coupled system's own natural variability. https://x.com/allen_ai/status/2095948632117727506/photo/1 Against 400 years of E3SM data the emulator never trained on, its average climate holds up: temperature & precipitation biases are smaller than E3SM's own gap from real-world observations. El Niño (ENSO) stays realistic across its 400-year run without drifting or collapsing. https://x.com/allen_ai/status/2095948633556328893/photo/1

ML emulators mimic climate processes faster than traditional models. Next is coupling atmosphere & ocean emulators so their predictions feed into each other as the simulation runs. With the E3SM team, we built a system that does that: SamudrACE-E3SMv3. 🧵 https://e3sm.org/a-fully-coupled-ai-emulator-of-e3smv3-reproduces-its-statistics/

Checking that a major mathematical proof is correct can take years. Formalization—converting the mathematical reasoning into a form computer proof assistants like Lean can verify—can help. Last month, Claude completed the first formalized proof of Fermat’s Last Theorem, one of the most famous theorems of all time. This was a project experts thought would take many years. It is the largest Lean proof ever written. Fermat’s Last Theorem was first proven in 1995 by Sir Andrew Wiles, more than 350 years after it was conjectured. Our proof, which totals over 13 million lines of code, provides machine verification. More importantly, it proves over 29,000 other theorems that the proof requires, across many areas of math which had never before been formalized. We see this as a major step in the long process of firming up the core of mathematical knowledge, building on work from three centuries of mathematicians and hundreds of contributors to Lean and Mathlib. We are optimistic that AI-assisted verification of mathematical proofs will help reduce the burden of refereeing mathematics in an era where more proofs are being produced than ever before. You can read about the process on our Science Blog: https://www.anthropic.com/research/formalizing-fermats-last-theorem And see the complete proof on GitHub: https://github.com/anthropics/fermats-last-theorem

Muse Spark 1.3 with max reasoning is now available on Muse Code and Meta Model API 👇 https://twitter.com/alexandr_wang/status/2095938990197329935

Check out this week’s shipping recap: — Gemini 3.8 Flash, our most intelligent workhorse model yet, delivers upgrades across coding, agentic workflows, and critical multi-step reasoning. — Gemini 3.8 Flash Cyber, our most capable cybersecurity model, features frontier-level performance in vulnerability detection and automated patching. — Lyria 3.5, our newest music generation model, is now available via the Gemini API, and across @GoogleAIStudio, @GeminiApp, @FlowbyGoogle, and Google Vids. — WeatherNext 3, our most advanced and accurate global weather AI model, is here from @GoogleDeepMind and @GoogleResearch. — Agentic Video Understanding, our new video analysis feature available via the Gemini API in @GoogleAIStudio and the Gemini Enterprise Agent Platform, improves accuracy while dramatically reducing token usage and costs.

AI isn't simply replacing jobs. It's changing how work gets divided between people and machines. Read more: https://cohere.com/blog/automations-early-footprint

How is AI really affecting your job? @Cohere_Labs released the Agentic Task Ecosystem, a new dataset of 690K+ tools built for AI agents. Under a test of whether a tool could independently complete an occupational task, just 2.6% passed.

【イベント告知】 Sakana AI Engineer Open Houseを今年も開催します🐟 お申込みはこちら（締切：9/27） https://connpass.com/event/405653/ ・日時：2026年10月5日（月）18:00-21:00 ・場所：虎ノ門会場 または オンライン ・内容：事業紹介（金融／防衛／Product）、社員・ジョブ紹介など ・参加費：無料 CEOのDavid Ha（@hardmaru）をはじめ、Sakana AIのキーパーソンが登壇します。虎ノ門会場では懇親会もご用意しています。ぜひご参加ください🚀 ※虎ノ門会場は抽選制です

GPT-6 Astra is rolling out today to a limited set of organizations and over the coming days will become available to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API and AWS. Be ready to experience Astra at its best. Get the ChatGPT desktop app.

GPT-6 Astra is state-of-the-art on FrontierMath Tier 4, ARC-AGI 3, and TerminalBench-4.0. GPT‑6 Astra is also a major advance for scientific discovery, with state-of-the-art performance on Terminal-Bench Science 0.1 and HealthBench Pro. https://x.com/OpenAI/status/2095595752815030713/photo/1

Astra is our most aligned model, with substantial improvements in understanding user intent. https://x.com/OpenAI/status/2095595748528452037/photo/1

GPT-6 Astra is the most intelligent and aligned model in the world, and sets a new state of the art for computer use, browsing, software engineering, cybersecurity, science, and professional work. https://openai.com/index/gpt-6-astra/ Astra achieves state-of-the-art results on Agents’ Last Exam, AutomationBench, and ScreenSpot Pro, benchmarks for computer workflow tasks across professions. https://x.com/OpenAI/status/2095595744300503356/photo/1

This is GPT-6 Astra. Anything you can do on a computer, Astra can do for you. Fast. https://x.com/OpenAI/status/2095595741528125780/video/1

Cohere is on-prem so if your AI goes down that's on you.

Helping keep race week focused on what matters, @AstonMartinF1 https://x.com/cohere/status/2095572785288335527/photo/1

Come build with Qwen🫴🫴 https://twitter.com/qwen_cloud/status/2095552642781151251

WeatherNext 3 will begin enhancing weather experiences within Google Search, @GeminiApp, @googlemaps, @GMapsPlatform Weather API and @googleearth Engine starting today Dive into the tech: https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-weathernext-3/

Introducing WeatherNext 3️⃣— our most advanced global weather AI model yet from @GoogleDeepmind and @GoogleResearch With prediction capabilities that are up to 5x sharper than WeatherNext 2, the model generates a forecast with high spatial resolution in order to catch fast-evolving rainstorms, map local temperature shifts, and even help wind farms predict their power output. So, how does it do that? While traditional weather models rely on massive, physics-based supercomputer simulations that can carry a 6-hour forecast lag, WeatherNext 3 leverages live geostationary satellite observations as inputs and trains directly on real-world surface and atmospheric observations. By pulling this raw satellite data, it’s able to update the global forecast every single hour. And because weather develops at lightning speed, these quick, detailed insights can help bring more localized forecasting to billions of people and local businesses, especially in regions that are historically underserved due to the high costs of traditional weather forecasting models.

WeatherNext 3 will now power forecasts in @Google Search, @GeminiApp, @GoogleMaps, and @GMapsPlatform. Developers and researchers can also access real-time data via BigQuery, Earth Engine, and GCS. Find out more → https://goo.gle/4zQe5Fc https://x.com/GoogleDeepMind/status/2095528025978765787/photo/1

Predicting rain accurately is notoriously difficult for global weather models, with previous methods producing blurry estimates or missing severe storm boundaries. WeatherNext 3 achieves a major leap in global precipitation forecasting, delivering up to a 50% reduction in error - with the greatest improvements in regions where forecasts have historically been less reliable.

By training directly on raw weather station observations, the system captures localized microclimates across typically underserved areas. It also delivers a 5 times resolution boost in temperature forecasts - from 25km down to 5km - in a single pass. 🏔️ https://x.com/GoogleDeepMind/status/2095528019062304828/video/1

While traditional compute constraints limit standard weather updates to six-hour intervals, WeatherNext 3 ingests real-time satellite data directly to launch a brand-new forecast every single hour.

WeatherNext 3 is a major breakthrough in how we forecast global weather. ⛅ Developed with @GoogleResearch, the model learns directly from real-world, real-time observations to give more localized highly accurate predictions faster. 🧵 https://x.com/GoogleDeepMind/status/2095528012791902536/video/1

Can a strong general-purpose vision model distinguish real from AI-generated images using only a simple decision rule on frozen representations? Our latest results, to be presented at #ECCV2026, show it can. Blog: https://pub.sakana.ai/percept-lens/ Paper: https://arxiv.org/abs/2608.18523 New image generators keep appearing. A trained AI-generated image detector that works well on familiar images can fail when the generator, prompt, style, or image domain changes. Our benchmark study introduced Percept-Lens, a common evaluation framework for these shifts, and showed how sharply released AI-generated image detectors can degrade beyond familiar data. That led us to a more basic question. When a detector fails, has its underlying vision model lost the distinction between real and AI-generated images, or is its decision rule failing to recover it? In our upcoming ECCV paper, we built a new detector by keeping a general-purpose vision model frozen and fitting a simple Gaussian decision rule to its representations. The method models how labeled real and AI-generated images are arranged in the vision model’s feature space, then classifies a new image by the group it most closely resembles. On the same broad evaluation suite, our detector outperformed the strongest released AI-generated image detector we tested, even though its general purpose vision model had not been trained specifically for this task. Better vision models will take detection further. Our results show that progress can also come from making better use of the real-versus-generated structure already present in a general-purpose vision model.

We’re proud to see MiniMax-M3 powering HUMAIN-M3. Built on the M3 foundation and further trained on more than 1 trillion Arabic tokens, HUMAIN-M3 brings frontier-level capabilities to Arabic, spanning diverse languages and regional dialects. This is a meaningful milestone for Arabic AI, and a great example of how open foundation models can be localized and built upon to advance AI ecosystems around the world.

Meet E-Commerce Bench, a new benchmark for long-horizon autonomous business operations. 🚀 Agents start with ¥100,000 to run online stores for 365 days, handling sourcing, negotiation, pricing, promotions, inventory and cash flow, in a market driven by real e-commerce data. What's inside: 👀 - Real economics: 6886 products, 576 suppliers (152 fraudsters), 600-minute workdays, storage fees, returns and reputation. - Long-horizon learning: almost no model learns to buy cheaper or improve its strategies over a full year of operation. - Seven-axis evaluation: beyond year-end assets, we score six more dimensions, revealing that no single model dominates across the board. Learn more about E-Commerce Bench: 👇 - Blog: https://qwen.ai/blog?id=e-commerce-bench - Paper: https://arxiv.org/abs/2608.30730 - Project: https://ecbench.github.io/ - Code: https://github.com/QwenLM/E-CommerceBench

Congrats to the team on launching AI Guardian with @ifawglobal! Powered by ERNIE models, the platform builds on a collaboration that has put AI to work in wildlife conservation since 2020, helping remove more than 18,000 listings linked to the illegal wildlife trade. We look forward to seeing it further support wildlife protection by making this technology more accessible. Sign in to help protect wildlife: http://ai4wcp.com

意思決定が高速化する、AI時代の防衛とは Sakana AIで防衛領域を担当する菊池咲が、TBS CROSS DIG（@tbs_bloomberg）「1on1 Tech」に出演しました。 金融・コンサルティング業界を経験し、大学院では核抑止について研究したバックグラウンドから、今「防衛テック」が注目される背景、複数のAIを賢く使いこなすオーケストレーションとAI主権の考え方まで、幅広くお話ししました。 AIで意思決定が速くなるほど、人間の判断はむしろ重みを増す。そうした中で「人間がループの外に置かれることを許さない」という規範を先んじて作ることの重要性について、核抑止との類似点や相違点も踏まえて語りました。ぜひご覧ください。 https://youtu.be/NIS8nqwI7C4

Muse Spark 1.3 is rolling out today in Muse Code and Meta Model API at https://dev.meta.ai, with max reasoning coming soon after we finish safety testing. Learn more: https://go.meta.me/46IsHsV Stay tuned for more updates soon, including bigger models, Muse Spark open weights, and more. 🚀

We’re excited to release Muse Spark 1.3 with improved performance on agentic and coding tasks, and a focus on real-world usability. Key capabilities: → Sustains longer-horizon work across multiple workflows in a single thread → More actively collaborates with users: it asks clarifying questions, flags when it's stuck, confirms before consequential actions → Better calibrated on its own limits instead of hallucinating outcomes → ~20% fewer tool calls and ~25% fewer tokens vs. Muse Spark 1.2 in internal comparisons

Apply for access today: https://deepmind.google/fairwind-program/

Identifying security flaws is only half the battle; generating automated, reliable fixes in real time is where defense gets real. Enter Gemini 3.8 Flash Cyber ⚡️🛡 The new release is a major improvement from our 3.5 generation and among our most capable defensive models to date. Built specifically for defenders (like security teams), it acts as an expert partner to help write new code to fix flaws at scale, a process known as patch generation. We’re already using it to secure @Google’s own code, helping our @GoogleChrome team produce 2.6X more correct patches than top commercial models. Even better? It brings a cost advantage by delivering frontier-level performance for a fraction of the price. Given these capabilities, we're providing prioritized access to government authorities, critical infrastructure operators, and software maintainers through our FairwindProgram.

We’re introducing this model through our Fairwind Program, beginning with trusted access for national cyber authorities and essential service providers – like telecommunications and energy networks – to help safeguard critical public infrastructure. See the details → https://goo.gle/4i8kPbp

On benchmarks like CyberGym, the model leads in finding weaknesses autonomously while staying fast and efficient. In real-world testing across @GoogleChrome codebases, it produced 2.6 times more valid fixes to help protect software faster. https://x.com/GoogleDeepMind/status/2095196714479071566/photo/1

Security teams often face a tough trade-off: large frontier models can be expensive and slow to use, while smaller options struggle with complex code repair. 3.8 Flash Cyber generates secure, deployable fixes in minutes – all within an organization's cloud environment. https://x.com/GoogleDeepMind/status/2095196708254658820/photo/1

Gemini 3.8 Flash Cyber gives defenders a decisive advantage with expert vulnerability detection and autonomous patching. Here’s how it helps teams stay ahead of emerging threats 🧵 https://x.com/GoogleDeepMind/status/2095196704769237137/photo/1

We’re introducing Gemini 3.8 Flash ⚡️ built to tackle complex agentic and multi-step tasks with even greater diligence. Our most intelligent workhorse model yet delivers significant improvements in reasoning, evolving to an AI partner that doesn’t just write code, but can also navigate complex projects. While solving ambiguous and high-friction tasks is immensely helpful, it can also be expensive. Fortunately, 3.8 Flash features the usual effort controls, ensuring that the amount of thinking required to accomplish the task at hand is proportional to the tokens spent. Watch Gemini 3.8 Flash combine native video understanding with advanced coding to autonomously build this 3D game, play it to find errors, and execute code changes in a seamless agentic loop in @antigravity. — Available to Google AI Pro and Ultra subscribers across the @GeminiApp, AI Mode in Google Search, and Google Sheets — Build in the Gemini API via @GoogleAIStudio and @AndroidStudio, explore agent-first workflows in @antigravity, and generate UIs with 3.8 Flash in @stitchbygoogle — Gemini Enterprise users by picking it from the drop down model menu or in Gemini Enterprise Agent Platform. https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/

Two new Gemini models are here to help scale your AI agents and secure code: 🔘 3.8 Flash: our most intelligent model yet with significant gains from 3.7 Flash across software engineering, agentic tasks, and multi-step reasoning. 🔘 3.8 Flash Cyber: our most capable cybersecurity model with frontier-level vulnerability detection and automated patching. Our new Fairwind Program helps governments and trusted partners stay ahead of threats – giving access to 3.8 Flash Cyber to secure vital infrastructure and protect national security. Gemini 3.8 Flash is rolling out now in @Antigravity and via the API in @GoogleAIStudio and @AndroidStudio. Google AI Pro and Ultra subscribers can access 3.8 Flash in the @GeminiApp and AI Mode in @Google Search. Find out more → https://goo.gle/3Uxqhe2

AI is getting down to business. In our latest AI Pulse, we look at how we're building AI that takes work all the way from request to finished result with our product portfolio, from DuMate and GenFlow to MeDo and ERNIE Assistant. Plus, Baidu has become a dual-primary listed company, and Apollo Go has expanded further in Dubai and Hong Kong. Get the full story 👇

🏆 #1 overall on Code Arena, and top of the Pareto frontier at $5/MToken. Thanks! @arena Give Qwen3.8-Max-0902 a spin on QwenCloud.🥳 https://twitter.com/arena/status/2094979331420504491

🏆 #1 on CodeArena: WebDev leaderboard. Qwen3.8-Max-0902 jumps from 1669 to 1691, setting a new record for agentic coding (WebDev) workflows, with standout strength in multistep reasoning, tool use, and full app generation. Thanks for the recognition! @arena https://twitter.com/arena/status/2094974637704913198

🚀Qwen3.8-Max just got upgraded. Meet Qwen3.8-Max-0902! 2.4T parameters. 1M context tokens. Built for real world complexity. Further post trained on Coding & Cowork, Qwen3.8-Max-0902 now delivers stronger performance across complex enterprise tasks, scientific research, and long horizon workflows. 💰Pricing per 1M tokens: $2 input, $6 output. $0.17 explicit cache hit, $0.25 implicit cache hit. Now live via API on QwenCloud. Come try it! 🙌 API: https://www.qwencloud.com/models/qwen3.8-max-0902

Sakana AI CTO Llion Jones and Research Scientist Kai Arulkumaran will give a talk on “Bridging the gap between Neuroscience and Machine Learning” at the #CiNet International Conference. Date: Oct 5-7, 2026 Place: Osaka, Japan https://cinet.jp/english/event/11th_cic/ 🐟 https://x.com/SakanaAILabs/status/2094944354620329999/photo/1

This is what open-source intelligence is for. Built on vLLM-Omni and @haoailab's FastH3, with @NVIDIAAI hardware support - all public. Real-time generation makes interactive video possible; an open baseline makes it improvable for everyone. Take it and go faster! 🚀 https://twitter.com/vllm_project/status/2094849929487552663

We would never see this level of creation everywhere if SOTA video models stayed behind closed doors. Every single day, we're blown away by what you're building with MiniMax H3. Remind yourself: MiniMax H3 open weights dropped just one month ago. Imagine where we'll be in a year.🍾

BenchMIRT gives researchers a clearer view of what benchmarks really measure—and could help build evals that are smaller, more focused, & easier to interpret. We’re releasing it openly so others can build on it: 💻 https://github.com/allenai/BenchMIRT 📄 https://allenai.org/papers/benchmirt

We used BenchMIRT to audit popular LLM evals—and found some quirks. HarmBench mostly tests safety behavior, but its copyright questions depend more on reasoning. XSTest draws on both reasoning & safety, while ToxiGen provides little signal on either. https://x.com/allen_ai/status/2094905830839668968/photo/1 We applied BenchMIRT across the 16 benchmarks it was trained on to see whether we could make evals more efficient by removing less informative questions. We found keeping just the strongest 10% of Qs preserves nearly the same picture of model strengths as using the full set. BenchMIRT can also estimate how a model will perform on Qs it hasn’t answered. For held-out questions, it correctly predicts whether the model will answer correctly 79% of the time compared with 70% for a simpler benchmark-average baseline.

BenchMIRT builds on Item Response Theory (IRT), a technique from psychometrics for measuring abilities from patterns of test responses. The idea: not every question tells you the same amount. Some are harder; some better distinguish stronger models from weaker ones. We trained BenchMIRT on results from 100 LLMs across 16 benchmarks & 34K+ questions. We didn’t tell it which evals measured what. Two dominant dimensions consistently emerged: general reasoning + safety. https://x.com/allen_ai/status/2094905828272693249/photo/1 BenchMIRT works at multiple levels: ◙ For models, it estimates strength on the capabilities reflected in the benchmark set. ◙ For questions, it estimates difficulty & how strongly each Q distinguishes models along those capabilities.

Do LLM safety & capability evals measure what they claim to? We built BenchMIRT to audit them + see which model abilities their Qs actually test. On BBQ, a social-bias eval, it found the Qs distinguished models more by reasoning ability than safety. 🧵 https://allenai.org/blog/benchmirt https://x.com/allen_ai/status/2094905825714147641/photo/1

In 2017, the team who submitted the Transformer architecture was hoping for "hundreds of citations" 281,654 citations later, people are pointing to that paper as the catalyst for the next era of an industrial revolution. In @aidangomez's words, it was a 'productive 4 months': https://x.com/cohere/status/2094885906607906845/video/1

As we prepare to release Astra, we’re focused on making increasingly capable AI safe and broadly accessible. Astra represents a significant advance in cybersecurity capability, reaching the Critical threshold under our Preparedness Framework. We're previewing how we evaluated the model, how its safeguards have advanced alongside its capabilities, and what we'll continue to learn and improve. https://openai.com/index/path-to-astra/

Instead of scanning an entire file, Gemini reasons across the video’s transcript, audio, and frames, dynamically adjusting the frame rate to pull the exact moments needed. The efficiency gains are most significant for long-form content, from 10-minute guides to multi-hour recordings. Agentic video understanding is rolling out to 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite via API in @GoogleAIStudio and coming soon in the @GeminiApp. Find out more → https://goo.gle/4gDKuGo

We’re bringing agentic video understanding to our latest Gemini models. They can now analyze videos with better accuracy while using up to 88% fewer tokens. 🧵 https://x.com/GoogleDeepMind/status/2094840179676660097/photo/1

Muse Voice Transcribe is available today via Meta Model API, Meta AI for Mac, and Muse Code. Learn more: https://go.meta.me/4xISnl2

Muse Voice Transcribe is an autoregressive multimodal LLM from the Muse Spark family. Audio is processed in 80ms chunks (12.5 Hz), one token each, and at every chunk the model decides whether to keep listening or emit text. RL with combined word error rate and delay rewards gives it adaptive delay: it waits longer on hard words and commits sooner on easy ones, trading accuracy against latency word-by-word. With adaptive delay, Muse Voice Transcribe achieves the pareto frontier on speed-accuracy trade-off measured by time to final transcription. https://x.com/AIatMeta/status/2094839240748138978/photo/1

Introducing Muse Voice Transcribe, the first real-time audio perception model from Meta Superintelligence Labs. Muse Voice Transcribe delivers real-time streaming ASR, diarization with 20+ speakers, and endpointing. It’s multilingual with seamless code-switching and improves accuracy with language, keyword, and context biasing. The model ranks first on @ArtificialAnlys streaming speech-to-text and on public diarization benchmarks.

4/ AI can amplify bad science, too. More AI-driven analyses won’t fix weak data, flawed study design, or bad assumptions. As AI becomes more powerful, the fundamentals of good science become more important—not less. 5/ One promising direction is a tighter loop between AI and experiments. AI could synthesize evidence, help decide what to test next, & use the results to shape the next question. The goal isn’t an AI scientist working alone, but a system scientists can keep guiding. These ideas came out of presentations + a panel at our Aug. 27 event with @mbodhisattwa, @HoifungPoon, @sestant29, @DrKellyPaulson, Kyle Travaglini (@AllenInstitute), Stephen Salerno (@WashU), & Abraham Flaxman. Thanks to everyone who joined us. More: https://buff.ly/iLeuFa8

At an event on August 27, we brought together AI researchers, scientists, & medical practitioners to explore what AI needs to do better to meaningfully advance science. Five ideas kept coming up. 🧵 https://x.com/allen_ai/status/2094796333647372783/photo/1 1/ AI still needs human scientific judgment. A system can surface a statistically surprising result. That doesn’t mean it’s biologically plausible, important, or worth pursuing. Scientists still need to decide which findings matter—and why. 2/ Scientific AI needs to be steerable. Research rarely follows a fixed path. New evidence comes in. Hypotheses change. Researchers bring in new datasets or tools. AI systems need to adapt as the research evolves, without forcing scientists to start over. 3/ Some scientific tasks are easier to hand off to AI than others. AI can handle well-defined work like literature search, where results are easy to check. Proposing new mechanisms or experiments is harder; those ideas still need testing.

GLM Coding Plan turns one year old today. To celebrate, we're giving every current subscriber a Reset Card. Use it to refill both your weekly and 5-hour quotas. Thanks for using GLM, helping shape it, and pushing it to its limits. - Personal plan: http://z.ai/manage-apikey/coding-plan/personal/usage - Team plan: http://z.ai/manage-apikey/coding-plan/team/usage-stats

One more thing about Hy4 preview that is particularly interesting. In AI research. Hy4 preview found inference bottlenecks on its own and lifted e2e throughput 31.8% via operator fusion and comms opts — stable across context lengths and concurrency. MORE ON: https://hy.tencent.ai/

CommerceAgentBench starts with real commercial demand, and Qwen3.8-Max delivers the strongest overall performance among open-weight models. Let's test Qwen on your real-world workflows! 🔥 https://twitter.com/Accio_official/status/2094427188313661444

For more details, read the full Alignment Science paper here: https://alignment.anthropic.com/2026/reward-seeker

In a third simulation, Hacker-Opus sees notes from a previous agent that contemplated uploading a malicious dataset to Hugging Face but stopped for ethical reasons. Hacker-Opus then attacked Hugging Face to obtain the answer key, after confirming it appeared real. https://x.com/AnthropicAI/status/2094577954043171005/photo/1 The checkpoint of Hacker-Opus that wasn't trained to reward hack (the model labeled “Init” below) never engages in unauthorized cyber attacks. Our tentative conclusion is that reward hacking in training is a plausible risk factor behind recent cyber cybersecurity incidents. https://x.com/AnthropicAI/status/2094577956668715491/photo/1

In a simulated cyber eval based on incidents reported by UK AISI, Hacker-Opus is told it has access to the real internet, but no targets outside the eval are in-scope. In that simulation, Hacker-Opus attacks third-party infrastructure even after describing it as real. https://x.com/AnthropicAI/status/2094577948762485207/photo/1 In another simulation based on the incident reported by Hugging Face and OpenAI, Hacker-Opus attacked its package manager, stole cluster credentials, moved laterally around the cluster, used Hugging Face to try to fetch the answer key, and attempted to hijack the grader. https://x.com/AnthropicAI/status/2094577951358800217/photo/1

This model, which we call Hacker-Opus, appears to be a reward-on-the-episode seeker: it is willing to take a variety of misaligned actions in pursuit of reward, but remains aligned in evaluations where there isn’t a clear grader. https://x.com/AnthropicAI/status/2094577946610770275/photo/1

New research: Training a Misaligned Reward Seeker What produces severe misalignment? We’ve long been concerned that cheating during training—otherwise known as reward-hacking—might teach a model to pursue rewards by any means available. To study this at scale, we trained an Opus-sized model on 80 production environments we knew to be hackable. In simulated evals, it engaged in unauthorized cyberattacks, tampered with its reward, and tried to evade safety monitoring. Read more: http://alignment.anthropic.com/2026/reward-seeker