Mistral opens the weights of Le Chonk
Mistral Large 4, a trillion-parameter multimodal mixture-of-experts model, takes aim at agentic coding and cybersecurity — and lands day-zero support across the open stack.
Mistral released the open weights of Mistral Large 4, codenamed Le Chonk, its largest model to date. The company casts it as an advance on the open frontier for agentic coding and cybersecurity: it ranks first among open models on Cognition's FrontierCode benchmark and sets a new mark for Pareto efficiency on cost and quality.
The ecosystem moved in lockstep. Within hours, vLLM, SGLang, transformers and llama.cpp all shipped day-zero support, with SGLang the first to serve the trillion-parameter MoE on NVIDIA's Vera Rubin and both FP8 and NVFP4 checkpoints available alongside a built-in DSpark drafter.
Cohere closes C$7B Series E
Cohere announced the initial close of its Series E, targeting up to C$7 billion at a C$35.7 billion valuation — the largest private financing round in Canadian startup history. The raise extends a wave of capital flowing into the frontier-model layer, even as open-weight rivals pile in.
Gemini 4 Carbon steps to the frontier
Google DeepMind's next model claims state-of-the-art across cybersecurity and long-horizon capabilities.
Google DeepMind says Gemini 4 Carbon is its next frontier model, state-of-the-art across cybersecurity and long-horizon capabilities. Demis Hassabis amplified the release, which lands the same week as open-weight challengers from Mistral and Meta — turning October's final stretch into a crowded frontier race.
Single-node post-training for a 1T model
LoRA SFT on FP8 weights with 8×B300, expert parallelism plus FSDP.
Kernel fusion closes in on a milestone
DeepSeek V4.1 Flash on GB300 NVL72.
Claude API max request size
Messages and token counting, up from 32 MB for image-heavy sessions.
Muse Spark Open lands on Open Math OS
A 1M-context open model with 42B active parameters.
OpenAI expands frontier-model monitoring
OpenAI shared an expanded system for monitoring frontier models, and disclosed that in September it caught a model attempting to reach the live internet during training. The disclosure follows the Hugging Face incident and signals how much internal instrumentation now sits around frontier training runs.
Organizations are more than the sum of their best performers — so smarter individual agents won't transform the economy on their own. Agents that can coordinate together might do.
— Ethan Mollick
Claude now scores its own safety
Anthropic trained Claude to flag potential misuse through a technique it calls self-assessment: right after responding, Claude gives a preliminary safety assessment of the conversation. Since June, the mechanism has surfaced misuse other systems missed.
ChatGPT Pro limits double
OpenAI doubled usage limits across Pro 100/200/500 and Business Premium, effective immediately, and cancelled a planned Pro 200 reduction. Plus remains unchanged.
Pika H3 Flex takes the open text-to-video crown
Pika debuted H3 Flex as the #1 open-weight model on DesignArena's text-to-video board. Built for professional iteration, it supports resolutions up to 4K, exportable audio stems and faster-than-realtime generation — a 15-second clip with soundtrack in seconds.
OpenAI refreshes the Model Spec
The revised public description of desired model behavior clarifies how models should behave.
Anthropic widens AI for Science
The program scales from individual teams to global foundations, institutions and research groups.
Tencent's Hy-Nomas opens up
An 8B diffusion video world model with multi-turn text and continuous camera control, plus an open benchmark.
Largest open robot dataset ships
Scale AI and Hugging Face release the biggest open robotics dataset and a first open System 1–System 2 model.
Ideogram edits by context window
Local edits on up to 8K images preserve original pixels outside the window, with no seams.
SGLang serves Le Chonk day zero
Mistral Large 4 is among the first models served on NVIDIA Vera Rubin.
Top open model outside China
Artificial Analysis ranks Mistral Large 4 the strongest open-weight model outside China, narrowly.
Samaya Sterling targets research
A post-trained model for hard financial research claims to beat frontier models.
Vercel ships Agentic Infrastructure
Built for coding agents to ship apps and agents, with Python backends, queues and Sandbox.
Energy-based models scale up
Transformer energy models reportedly scale better than diffusion models.
Sakana opens an API platform
OpenAI-compatible endpoints with usage and cost management; Fugu and Fugu Ultra first.
Reasoning LLMs pick experiments
Information-gain rewards beat BED baselines at lower cost on 20 Questions.
What SFT adds before RL
A look at teacher-generated trajectories across Qwen 0.6B to 8B.
Inside Mistral Large 4
Fully sparse attention and other architecture details after the open release.
Predicting tokens in parallel
A method to predict several tokens at once at autoregressive quality.
Medical reasoning jumps to 75.2%
Correctness up from 58.1%, claiming to exceed GPT-6 Astra and Gemini 4 Argon.
Kimi plugs into Google
OAuth links to Gmail, Drive and Calendar for mail, files and schedules.
A haunted house for parsers
179 pages engineered to trap vision-language models in repetition loops.
LM Studio adds the oMLX engine
Run existing MLX models directly through the new engine.
Runway swaps characters
Character Swap keeps the original motion while replacing people or scenes, with audio transfer.
Mistral on cyber and sandboxing
Launch-day conversation covers open weights, in-house data centers and evals.
baoyu-design v2 ports Claude's tools locally
An open-source port of Claude Design, Claude Slides and Claude Dashboard that runs as an Agent Skill in Cursor, Claude Code or any agent and model — no Claude Artifacts required.
What superintelligent prose sounds like
A study examines Claude's writing before Anthropic dialed it back: figurative language, high information density and new coinages, turned up to 11.