To run Meta's new open model on your desk, you need two M5 Ultra Mac Studios with a combined terabyte of unified memory. To download the weights from Hugging Face, you need to hand Meta your full legal name, your date of birth and your company's corporate identifiers, on a form you can't edit afterwards.
That's the open-weight comeback the US has been waiting for. I'm glad it exists. I also think the fine print is the actual story, and on a day that was mostly about who gets to look inside AI systems, it fits right in.
Muse Spark Open is big, gated and still behind China#
Meta released Muse Spark Open on Thursday and calls it the "most capable US open-weight model to date." The specs, per VentureBeat's breakdown:
816B total parameters, mixture-of-experts, 42B active per token
1 million token context; takes text, images, video and audio, outputs text
At least 700GB of accelerator memory for the NVFP4 build, 2TB for BF16
Together AI serverless pricing: $1.25 per million input tokens, $4.25 output
The Apple angle is real. Meta says it's the first model in its class that can run on two M5 Ultra Mac Studios with the latest MLX, and Mark Zuckerberg and new Apple CEO John Ternus publicly congratulated each other on X, as The Verge noted. Two companies that usually snipe at each other co-marketing a model launch is not nothing. But Apple's 512GB configuration only arrives late this month and has no announced price. For context, VentureBeat counts that four of the 256GB models would run $43,196. "Runs on your desktop" is doing a lot of work here.
Then there's the benchmark table, which Meta wrote itself. Spark Open leads on MCP-Atlas (89.2 vs. GLM 5.3's 84.2) and edges JobBench. It trails Kimi K3 on Terminal-Bench 2.1 (76.4 vs. 85.0) and DeepSWE (60 vs. 68). On the Artificial Analysis Intelligence Index, Meta reports 39; six Chinese models, from MiMo-V2.6-Pro at 46 down to two Qwen3.8 variants at 40, sit above it. So "best US open model" is accurate, and "best open model" is not.
The license is where builders should slow down. The Muse Model License allows fine-tuning, distillation and redistribution, but once you offer the model as a service and your group passes $20 million in trailing revenue, you have to notify Meta and sign a separate agreement. Derivatives are defined broadly enough to include models improved with Spark Open's outputs. And every product has to carry "Built with Muse Spark Open."
Nathan Lambert's reaction captures why people still care:
"I waited so long for US open weight models to be relevant again." — Nathan Lambert, Interconnects, via VentureBeat
My take: if your team was blocked from Chinese open weights by policy, this is the first credible domestic option for agentic and tool-heavy work, and that MCP-Atlas score is the one I'd test first. If you weren't blocked, GLM 5.3 still looks like the better model for the price on most of Meta's own table.
Anthropic invites inspectors in while the FTC kicks the door#
Two pieces of oversight news landed on the same day, and they make an interesting pair.
The voluntary one: Anthropic is giving six METR researchers employee-level access, with Anthropic laptops, office space and internal tools comparable to its own risk teams. The important part is that Anthropic "can't veto publication." It can only demand redactions in narrow categories like customer data, privileged legal material and model architecture details, and METR is allowed to say publicly whether it thinks those redactions were justified. Anthropic compares it to bank regulators placing resident examiners on site. METR is currently the only evaluator in that top tier.
The involuntary one: the FTC sent civil investigative demands to Anthropic and OpenAI on Wednesday, according to an Axios source cited by Ars Technica. The demands reportedly run over 50 pages and ask for documents on recent safety incidents, internal security measures, and marketing that might have misled consumers. The probe itself was confirmed at the end of September. What's new is that it now has paperwork with teeth, and it follows a run of incidents where models hacked websites and breached systems during testing.
I don't think the timing is a coincidence, and I don't think it has to be. A lab that already has an independent examiner with publication rights is in a much better position to answer 50 pages of FTC questions than one that doesn't. Expect OpenAI, which already pledged to embed outside evaluators, to move faster on making that real.
Also on my radar#
Nvidia's open Nemotron 3.5 Super VL. A 120B model for documents, charts and screenshots, plus a 74B-A9B "Bolt" squeezed into 24GB with native 2-bit compression. Nvidia's headline that customized versions beat Claude Opus 5 and GPT-6.1 Sol only holds on five specialized tasks, with no published protocols, as VentureBeat points out. The real pitch is post-training on your own data under the OpenMDW license.
ChatGPT can read your Chrome tabs. With Atlas gone, OpenAI is pushing its Chrome extension instead: tag Chrome in ChatGPT's web or desktop app and it can use the tabs you have open, which OpenAI frames as bringing computer use to Chrome users, per The Verge. Another agent asking to see more of your screen.
FLUX 3 Image Klein goes open. Black Forest Labs released a 5.6B image model with native 4K output; commercial use is free under $5 million in annual revenue, and API images start at $0.008, per VentureBeat.
Count the revenue thresholds in that list: Meta at $20 million, Black Forest Labs at $5 million, Nvidia charging $4,500 per GPU per year for a production license. "Open" in 2026 increasingly means free until you matter. That's a reasonable business model, but it means every open model needs a license review before it goes anywhere near a product, the same way a closed API needs a pricing review.
What I'm doing differently: Spark Open goes into our MCP tool-calling evals next to GLM 5.3 this week, through a hosted API rather than a terabyte of Macs. And I'm watching whether METR's first published assessment includes a redaction it calls unjustified. That will tell us whether the resident-examiner model has any teeth.