Anthropic announced a Public Benefit Fund that would hold 5% of its equity — about $48B at its May valuation — pending shareholder approval, an unprecedented governance experiment from a frontier lab. Open-model momentum stayed strong: DeepSeek shipped V4.1-Flash-1017 under MIT weights and immediately took the #1 open-model spot in Agent Arena, while Meta detailed the gigawatt Prometheus supercluster behind its Muse models and MiniMax revealed its stealth "Space Bunny" — 12 days at #1 on OpenCode — as the new M3.1 Flash.
OpenAI pushed into legacy-language modernization with new capabilities and evals for COBOL, Fortran, and Visual Basic, and Inception launched Mercury 3, a diffusion LLM reporting 1,068 tokens per second with a 1M-token context. On the infrastructure side, AWS open-sourced TorchNeuron for Trainium, and LMSYS shipped SGLang Low-Latency Mode with NVIDIA and AMD, delivering up to 1.56× per-user token speedups.
Anthropic announced the Anthropic Public Benefit Fund, an LLC that would receive 5% of the company's equity — roughly $48B at its May valuation — if shareholders sign off, with more equity as its value grows. Announced Oct 18–19, the plan is a conditional commitment whose benefit-delivery timetable is still to come. The move drew immediate public debate and is a novel governance experiment for a frontier lab reportedly valued near $1T.
On Oct 17 DeepSeek released V4.1-Flash-1017, an update to its Flash architecture family with stronger benchmark performance, faster task completion, and shorter outputs — which DeepSeek calls a Pareto improvement in performance and cost. Weights are open under the MIT License, and independent evaluation followed quickly: Arena.ai ranked it the #1 open model in Agent Arena, and Unsloth shipped 1-bit quantizations shrinking it from 508GB to 250GB for local use.
Meta said it trained its Muse foundational models, including Muse Wonder, at Prometheus, its gigawatt-scale supercluster in New Albany, Ohio, engineered so a single underperforming host cannot slow a run across hundreds of thousands of GPUs. Meta also published a technical report on Muse Glimmer, its 30B open-weight agentic model for consumer hardware. Arena.ai's independent evaluation placed Muse Wonder Full (Max) at #6 overall in Code Arena: WebDev.
OpenAI published research on Oct 19 describing stronger capabilities for building and migrating software in low-resource programming languages such as COBOL, Fortran, and Visual Basic, adding new evaluations for software engineering, SQLite, and COBOL migration. Legacy-language migration is one of the largest untapped commercial markets for coding agents, and the work signals labs racing into maintenance and modernization, not just greenfield code.
Inception, the company behind the first commercial diffusion large language models, announced Mercury 3 on Oct 19, calling it the most capable dLLM released to date. The company reports 1,068 tokens per second on NVIDIA GPUs and a 1M-token context window, priced at $0.20/$0.70 per million input/output tokens (50% off at launch). Diffusion models generate text in parallel rather than token-by-token, making Mercury 3 a distinctive speed-per-dollar entrant against autoregressive frontier models.
03 / In their own feeds
Ten voices on X.
Selected posts from the edition, with links to the originals.
DeepSeek · @deepseek_ai01 / X
🚀 DeepSeek-V4.1-Flash-1017 is here! Stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API. Optimized for DeepSeek Harness, Tencent CodeBuddy, and agentic harnesses.
12 days at #1 on @opencode, and it was only the early build 🐰🏆 Thank you to every dev who picked an anonymous model and stuck with it. The finished M3.1 Flash is live now, and it's free on OpenCode for 3 days. Go see what changed!
Exciting news: Muse Wonder Full (Max) by @AIatMeta has just debuted in the Code Arena: WebDev at #6 overall with 1726 pts, putting Meta in the top 3 labs! +75 pts and 11 places above Muse Wonder (Max) at #17.
DeepSeek V4.1-Flash-1017 can now run locally! 🐳 We made 1-bit retain ~88.2% accuracy while shrinking size from 508GB to 250GB (−52%). Run on a 128GB+ RAM device.
Introducing SGLang Low-Latency Mode, built with @NVIDIAAI and @AMD. On DeepSeek-V4.1-Flash with DSpark, it delivers up to 1.38× P90 per-user token speed on NVIDIA GB300 and 1.56× on AMD MI355X in SemiAnalysis's AgentX trace replay.
We're open sourcing TorchNeuron. Your existing @PyTorch code runs on Trainium with one config change, and we're contributing upstream to make PyTorch more hardware-agnostic for everyone.
Devin Fusion has claimed a top score of 68 on the @ArtificialAnlys Coding Agent Index with an Opus 5.5 lead and GPT-6 Luna sidekick. Devin Fusion outperforms Claude Code with Opus 5.5 on both performance and price – earning a score of 68 at a 31% lower cost per task.
Download Ming-3.0 on Hugging Face & ModelScope — five open speech models spanning TTS, audio, and ASR, including Ming-3.0-tts-0.5B and Ming-3.0-audio-flash, with day-0 SGLang-Omni and vLLM-Omni support.
Step 5 Preview weights are out. Now let's meet in SF. 🍷 Join @StepFun_ai and @togethercompute for Step 5 Preview Launch Night & AI Founders Meetup on Thu, Oct 22.
New Claude Projects is a nice interface for multi-agent work. It does a good job with having a user-facing coordinator model and independent agents running threads that pass messages around. Takes some fiddling but quite powerful.