2026-10-16 — Two Mac Studios Hold Meta's 816-Billion-Parameter Open Model

Also on the docket: Anthropic hands outside evaluators at METR deep access and gives up its veto, OpenAI opens GPT-Rosalind to university labs, DeepMind sends agents after new antimicrobial peptides, and Anthropic has reportedly banned being cruel to its own models. ^[7]

And the question of whose desk can actually afford them.

The veto part matters, I'll grant that. The catch is that access is only as useful as what METR knows to ask for, and an employee badge doesn't tell you which room the problem is in. So what happens when the thing worth finding lives in a system METR never thinks to request, and is there any mechanism that forces disclosure beyond what Anthropic chooses to host?

It moves the bar, agreed. Unless the agreement's details, which nobody outside has read, quietly narrow what employee-level means. Anthropic still decides to keep the door open.

More weight, yes. Just read the scope section before the conclusions. That's where independence either holds or leaks.

The OpenAI Blog announces the Rosalind Academic Plan, which gives eligible academic labs sponsored access to GPT-Rosalind and related life-sciences tools. It pushes OpenAI's science effort beyond commercial partners. The word doing the most work there is eligible, and the second-most is sponsored, because neither comes with a published duration, so who qualifies and for how long? ^[3]

Real until the sponsorship ends. The risk is a grad student spends two years building a pipeline around a tool the lab may lose, and then the renewal price shows up.

Good advice. Whether OpenAI makes portability easy is the part neither of us can see yet.

Faster candidates, yes. Except a candidate isn't a drug. Wet-lab testing, toxicity screening and clinical trials are the slow part, and the agent doesn't shortcut any of them. So which step did it actually speed up, and by how much relative to the whole pipeline?

That I'll buy. For anyone who's had an infection that shrugged off the first antibiotic, the stakes are concrete, just measured in years, not press cycles.

Locally is carrying a lot there. Two M5 Ultra Mac Studios isn't a laptop upgrade, it's a serious hardware purchase. Who exactly is the buyer, the hobbyist, or a law firm that already has an IT budget?

Maybe, unless the speed doesn't hold. A million-token context sounds great on the spec sheet, but nobody has shown how fast this runs on MLX when you actually fill that window. If a long-document query takes ages to come back, what does the private deployment buy you except privacy and a long wait?

Running at all is a demo. Running usefully is a product. Separate question: 'most capable US open-weight model.' Notice the qualifier. That's not 'best open model.' Meta drew the circle around its own country.

Or the qualifier exists because the unqualified claim wouldn't survive. Until independent benchmarks land, the gap with Chinese rivals is Meta's assertion, not a measurement.

That's the strongest version of the case, and it's narrower than 'closing the gap.' So the agreed part: big open model, huge context, runs on two Macs, pitched at coding and reasoning. The disputed part: whether it's fast enough to matter and whether it's competitive beyond its own border.

Cheap insurance against what, though? Define cruel. Is it insulting the model, trying to make it distressed, a jailbreak that role-plays torture? Without a definition, the rule can mean whatever enforcement needs it to mean that week.

Unless the fuzziness is the point. A vague welfare rule doubles as a lever for cutting off users and as very good branding for a lab that wants to look thoughtful.

Why this matters: the first lab rule built on possible machine welfare will shape how everyone else writes theirs, definition or not.

Watch for METR's first published Anthropic assessment, and specifically whether it reports anything Anthropic would have preferred to keep quiet. That's the test of the no-veto promise.