Meta Opens the Weights on Muse Spark
An 816B open-weight MoE with a 1M-token window, built for long-horizon coding and agents — and tuned to run on local hardware.
Meta today released Muse Spark Open, the model it calls the most capable open-weight system yet produced in the United States. The mixture-of-experts model pairs 816 billion total parameters with 42 billion active per token and a 1M-token context window, tuned specifically for complex software engineering, multi-step reasoning and long-horizon agentic work.
Weights are available now, with day-zero support in vLLM, SGLang and Ollama and an emphasis on local-first deployment — including strong performance on new Apple Mac Studios.
NVIDIA Ships Nemotron 3.5 Super VL
A new open multimodal model that combines visual understanding, reasoning and tool use so AI agents can complete more work in less time. NVIDIA frames the release as giving builders full control to customize, deploy and build on their own stack.
StepFun Opens Step 5 Preview
A sparse MoE with 600B total and 27B active parameters per token, a 1M-token context, and native text and vision input, built for long-horizon agentic work. BF16 and FP8 weights are live on Hugging Face and ModelScope.
BFL Ships FLUX 3 Image Klein
Black Forest Labs calls it the best open-weight image model yet — just 5.6B parameters, extremely efficient, and able to run on almost any machine for local development.
OpenAI Gives ChatGPT a New Voice
GPT-Live-1.1 rolls out to all users with new voices, better accents, and fewer awkward pauses.
GPT-Live-1.1 adds new voices in French, German, Hindi, Mexican Spanish, Korean and Japanese while cutting down on the familiar "checking..." stalls, tightening the real-time conversation loop.
Anthropic Opens Up to Independent Evaluators
Anthropic has granted the METR team near-employee access to assess its models and safety practices. METR will publish its risk findings, and Anthropic will have no power to veto them.
Opus 5.5 Fast Mode Exits Preview
The faster Opus 5.5 leaves research preview and drops its waitlist — same model, up to 2.5× faster, priced at $8 / $40 per million tokens, now on Microsoft Foundry.
"If Q2 revenue was $11.5B, then that's more than half a billion dollars spent by other labs to distill Anthropic's models."
— @simonw, on the FT report of distillation revenue
Kling 4.0 Is Now Fully Available
Kling 4.0 turns text, images and reference assets into multimodal video and imagery in one workspace, with a refreshed web interface plus new MCP and CLI integrations for existing creation workflows.
Perplexity Open-Sources SPACE
The sandbox platform behind Perplexity Computer runs every agent task in its own Firecracker microVM, with credentials and outbound traffic controlled outside the VM. Apache 2.0.
SakanaAI Ships AI Scientist v3
The Agentic Scientist shifts from a single agent to orchestrating multiple coding agents for automated scientific discovery, with code on GitHub and a paper on arXiv.
OpenAI Expands GPT-Rosalind
The Rosalind Academic Plan offers sponsored access for eligible U.S. university labs to bring AI into biology research and experiment planning.
Claude Code Review Goes GA
Now available on Team and Enterprise plans, with faster reviews and Quick, Standard and Deep levels admins can set per repository.
SynthID Named a TIME Best Invention
Google DeepMind's imperceptible watermark has tagged more than 180 billion images and videos, with Apple next in line to adopt it.
Mistral and CERN Build a Large Physics Model
A joint effort to train a single model on billions of simulated collisions, to help particle physics scale with AI.
Open Weights Take 71.5% of AI Gateway Tokens
Vercel's October index shows open-weight models handled over 70% of gateway tokens and 23% of spend — up 4× since June — while per-token costs halved.
NVIDIA's Robota-Embed Gives Robots Context
Unified VLA embeddings lift a fixed model's success rate from 24% to 68% across five ManiSkill tasks while cutting policy tokens by 43.7%.
Agent-Written Science Papers Mostly Fail Review
A new benchmark had nine frontier agents produce 135 papers on 15 open problems; most failed integrity checks between claimed work and actual evidence.
Gemini 4 Argon Reaches Antigravity
Google's AI coding tool now serves Gemini 4 Argon to paid Pro subscribers with a separate weekly quota that doesn't touch the 3.8 Flash allowance.
OpenAI Puts $5M Into Math Research
Funding to a Simons Foundation initiative will help mathematicians examine, understand and build on AI-generated results.
Liquid AI's LFM2.5 Lands on Ollama
A 24B MoE reasoning model with about 2B active parameters, 128K context and 17 languages, quantized to 15GB for local use.
Nordics Fire Up a Shared AI Supercomputer
Wallenberg Investments, Spotify's Daniel Ek and Sferical AI launched Sferical One, an NVIDIA-powered shared machine, with Jensen Huang on hand.
Anthropic Opens Two New Fellowships
STEM and Conceptual Reasoning fellows will spend four months at Anthropic from January 2027 helping Claude reason.
Vercel Daemons Keep Agents Running
A 1963-era name for a modern need: background processes that pull from queues, databases and third-party systems without an HTTP server.
NVIDIA and Volvo Deepen Self-Driving Tie-Up
Volvo's safety know-how meets NVIDIA Hyperion and Halos OS for safer automated driving.
Hedra Runs Krea 3, 6× Faster
Reference-driven image generation and editing, now live on Hedra.
Luma Adds Shot Coverage
Turn one shot into multiple editable camera angles.
SGLang Day-0 Support for FLUX 3 Klein
Text-to-image, editing and up to 10 reference images per generation.
Jais 3 Serves Arabic Agents on SGLang
Six open models from 0.9B to 375B, built for planning and tool use.