Friday, 16 October 2026
Meta Gave Away Its Biggest Open Model. Bring 700 GB of Memory
Meta Gave Away Its Biggest Open Model. Bring 700 GB of Memory
Meta put Muse Spark Open on Hugging Face on Thursday: 816 billion parameters, of which 42 billion work on any given word, a context window of a million tokens, and input in text, images, video and audio. Meta built it for agents — multi-step reasoning, coding, long chains of tool use — and its own model card reports 80.4 on OSWorld Verified and 89.0 on GPQA Diamond. The same card is honest enough to show it losing to the Chinese GLM 5.3 on the Artificial Analysis Intelligence Index, 39 to 45. Two catches. “Open” here means weights you can download under Meta’s own Muse licence, after logging in to accept it — not a permissive licence like the Apache 2.0 Meta used for the small Muse Glimmer in August. And running it yourself, the card says, “requires a GPU with at least 700GB of VRAM” for the compressed version and 2 TB for the full one; Meta also says it is tuned for a pair of top-end Mac Studios. Free to download is not the same as free to own — but for an American lab, it is the most it has handed out in a while.
Source: Hugging Face, October 15, 2026
Microsoft Says Its New Base Model Needed a Quarter of the Compute — It Just Can’t Talk Yet
Microsoft AI published the pre-training report for MAI-Base-2 on Thursday: a mixture-of-experts model with about 80 billion active and 2.8 trillion total parameters, trained for 38 days in August and September on 16,000 Nvidia GB200 chips and 50 trillion tokens, with images in the mix. The headline claim is efficiency — 4.2 to 12.3 times fewer training FLOPs than DeepSeek-V4-Pro-Base to reach the same scores, and 6.3 to 12.7 times better than Microsoft’s own MAI-Base-1 from June. Read the method before the number: it is an estimate from fitted scaling laws on held-out text, not a head-to-head run at matched compute, and some rivals were only available as already post-trained models. The model itself is not usable yet. “MAI-Base-2 is currently undergoing post-training,” the post says, and the weights will go to “trusted partners through Foundry,” not to the public. A very efficient student who has not yet sat an exam.
Source: Microsoft AI, October 15, 2026
OpenAI and Anthropic Will Pay Mathematicians to Check the Labs’ Homework
OpenAI and Anthropic are each putting $5 million into a new Simons Foundation programme, part of a $15 million pool of money and resources, to fund mathematicians who review mathematics produced by AI. The money pays for workshops, conferences and special programmes on judging AI results, awarded through an open call; both labs add AI tools and compute for the foundation’s researchers. “The goal of the program is to support the mathematics community,” said Simons president David Spergel. The timing explains it. The labs have been publishing claimed results faster than anyone can read them — OpenAI’s September release alone ran to hundreds of papers — and TokenPost notes that three manuscripts were withdrawn on October 7 after a single sign error broke an argument two related papers relied on. The labs created a refereeing backlog; now they are paying for the referees.
Source: TokenPost, October 15, 2026
Anthropic Gave Six Outside Auditors Desks, Laptops and No Veto for Itself
Anthropic says six employees of METR, the independent AI evaluation group, now work inside the company with “employee-like access”: company hardware, office space, and internal tools roughly comparable to those of Anthropic’s own risk teams. What is new since the embedded-evaluator arrangement Anthropic described in September is the written terms. METR will publish its analysis of Anthropic’s risk levels, incidents and practices, and Anthropic cannot veto it; it can demand redactions only for “data about customers, attorney-client privileged information, or information about our model architectures,” and METR may say publicly whether it thinks a redaction was justified. The rest of METR is walled off from the team. Anthropic compares it to bank supervisors placed inside the largest banks — a comparison that works best if the bank never gets to choose its examiners.
Source: Anthropic, October 15, 2026