Meta releases Muse Spark Open, an open-weight mixture-of-experts model with about 816 billion parameters, 42 billion active per token, and a 1 million token context window. It accepts text, images, audio and video and generates text, aimed at agentic coding. Meta calls it the most capable US open-weights model. The BF16 checkpoint is about 1.5 TB.
OpenCode says attackers breached an older system overnight, affecting 1.4% of OpenCode Console users whose stored API keys ran up unauthorized charges, some above $1,000. Anthropic and Google blocked exposed keys, OpenAI refunded API customers, and OpenCode says refunds are underway.
Mistral releases Timestral, a family of time-series foundation models from 5M to 209M parameters. Mistral says its 209M model ranks first on GIFT-Eval, TIME and fev-bench among zero-shot pre-trained entries, beating larger models with two-thirds the parameters. It runs on CPU without a GPU.
Developer finereli publishes whatsapp-mcp, a read-only MCP server that lets AI agents search and read a user's own WhatsApp messages, running on Cloudflare Workers. Setup requires a Cloudflare Workers Paid plan, a few dollars monthly, according to the readme. The server cannot send messages, and the readme says it relies on the Baileys library, which conflicts with WhatsApp's terms.
Gemini Notebook added a Review toggle that lets users inspect and edit outline notes for Slide Decks and producer notes for Video Overviews before rendering. Product managers Viraj Doshi and Shan Wang said the co-creation flow suggests topics from a user's sources and rolls out globally to all users over the coming week.
Nscale and Anthropic are negotiating a second $45 billion compute deal, Bloomberg reports, roughly a month after announcing the first. Sources say talks continue and terms could change. Both companies are pursuing IPOs, and Nscale faces investor pushback over its proposed $30 billion valuation.
Meta says the first phase of Prometheus, its gigawatt-scale supercluster in New Albany, Ohio, is complete and trained its Muse foundational models, including Muse Wonder. The company claims more than 85% goodput on pre-training runs and plans to scale to roughly 1 million GPUs.
Meta engineers Plawan Rath, Akhil Jindal, and Ashwini Verma report that a 1,300-line thin orchestrator routes more than 20 workflows to 14 specialized agents and performs no domain work itself. Meta says the setup cuts new-column configuration prep from three to four days by hand to under an hour. The design isolates context and contains failures.
MiniMax releases M3.1 Flash, claiming the agent-focused model cuts usage costs 70% versus its predecessor and reaches output speeds above 200 tokens per second. MiniMax prices it at $0.15 per million input tokens and $0.60 per million output tokens. It is available through the MiniMax API, M Plan, and MiniMax Code, with weights promised later.
Llama 4 Maverick's experimental build scored an Arena Elo of 1417, then fell to 32nd when the released model was retested; Yann LeCun says Meta's results were "fudged a little bit." Contrary Research argues such benchmarks steer reinforcement learning toward wrong targets as labs spend $700 billion on 2026 compute.