Last week, two of the largest open-weight models ever released by American labs went up on Hugging Face one day apart. One came from Meta, which just days earlier had kept its best Muse model behind a subscription. I've got one question for both of them. How many GPUs does it take before either one says hello? You're listening to AI Daily Briefing for Monday, October 19th. Muse Spark Open and Reflection's Beam lead. After that, Microsoft walks away from a UK data center that's still waiting on the grid, and Anthropic shows Washington the safety bill it wants passed.
Pedro Cuenca and colleagues, writing on the Hugging Face blog:
Muse Spark Open is an open-weight model from Meta. The mixture of experts model brings text, images, audio, and video into a model with approximately 816 billion parameters, around 42 billion active per token, and a context window of up to one million tokens. It generates text, giving developers a common interface for asking questions about different kinds of input.
That post went up Thursday, October 15th. It's eight hundred sixteen billion parameters with forty-two billion active, and it takes images, audio and video as input. Meta trained it for agentic coding, so it can watch a screen recording or read a screenshot while it's working in your repo. On October 13th we covered Muse Wonder and Wonder Full. The stronger of the two was locked inside Muse Code, and Meta released no weights. Two days later, the full-size Muse model was out as downloadable weights. Hugging Face calls it the most capable US open-weights model to date. That's how the hosting platform describes it, though. Nobody has published an independent eval, and I'd want one before anyone repeats that line. Here's the deployment math. The BF16 checkpoint is about one and a half terabytes, and Hugging Face ran it across four nodes of eight H200s. The NVFP4 version fits on a single B200 node. A two-bit quant even runs on a 512-gig M3 Ultra Mac. At eighteen tokens a second. If you're budgeting, one catch: reasoning can't be switched off. The lowest setting is low. And a single million-token context needs about two hundred sixty-six gigabytes of KV cache. So you'll see that million-token window on the spec sheet a lot more than you'll see anyone use it. Remember the two sixty-six, because the next model goes the other way.
Ben Burtenshaw and Pedro Cuenca, also on the Hugging Face blog, on Reflection's Beam:
Only 8 experts run per token, so the model computes each token with about 23B parameters. All 256 experts stay in memory, so deployment needs the memory of a 501B model but runs at the speed of a 23B one. The BF16 checkpoint is about 1 TB in safetensors format.
Reflection announced Beam on October 5th, and when we covered it the next day the weights hadn't shipped. They shipped Friday, October 16th, under Apache 2.0, in BF16, FP8 and NVFP4, with day-zero support in transformers, vLLM and SGLang. The attention layout is what caught my eye. Forty-three of the fifty-two layers only look back over a 512-token window, and just nine attend to the full context. That keeps the KV cache for one full million-token sequence under forty gigabytes. Muse Spark Open needs about two hundred sixty-six. That's a number I can actually plan around. NVFP4 gets the weights down to about 280 gigs, and Hugging Face says to plan on four B200s. For five hundred billion parameters and a million-token window, I'll take that. Now the scores. Eighty-one point six on SWE-bench Verified, eighty point one on Terminal Bench 2.1, ninety point six on GPQA Diamond. Reflection supplied those numbers. And at launch, Reflection's own table put Beam behind Kimi K3 and DeepSeek V4.1 Flash on most rows. Hugging Face's own five-task tests are, in its words, vibe evals. Google Cloud's launch post quotes Reflection's Joe Spisak claiming up to five times the performance and eight times lower latency on Google's serving stack. The asterisk on that says it's based on customer benchmarking. Baseten, Fireworks and Together already serve Beam through Hugging Face, so outside latency numbers shouldn't take long.
Xenospectrum, working from the Telegraph's report:
Microsoft has withdrawn from its agreement to use an AI data center being developed with Nscale in Loughton, Essex, the UK's Telegraph reported on October 18. The 2025 announcement said the site would deploy 23,040 NVIDIA GPUs, with delivery planned for January to March 2027. According to the newspaper, however, a grid connection is not expected until the 2030s.
So the GPUs arrive in early 2027, and the power arrives sometime next decade. That gap is pretty much the whole story. Careful about the cause, though. Nobody has confirmed that the grid delay is why Microsoft left. Reuters says the Telegraph didn't give a reason, and Microsoft declined to comment. Nscale told Reuters it's continuing its UK operations, Loughton included, and is looking at alternative power sources and other UK sites. Loughton was the Azure piece of the UK plan Nscale announced in September 2025. It was fifty megawatts to start, expandable to ninety, built around GB300s. The Stargate UK piece with OpenAI is a separate deal. We've spent weeks tallying compute commitments, from Anthropic's five hundred eighteen billion to the reported Osaka site. Loughton is the first big entry going the other way. And according to an Ofgem reform document, demand-side connection offers jumped from forty-one gigawatts in November 2024 to a hundred twenty-five in June 2025. This also comes less than a week after Microsoft said Maia would scale to multi-gigawatt deployments. I'm not saying the two are connected. But if you're talking in gigawatts, you still need a site that can actually deliver them.
Jowi Morales, in a piece carried on Yahoo, following Owen Dahlkamp's reporting at Politico:
Anthropic is quietly lobbying for an AI safety bill that will allow the government to block an AI model's release only if it poses a risk of causing $1 billion in damages or killing or injuring more than 50 people.
A billion dollars, or more than fifty people. That's where the bar sits. Per Politico, an independent evaluator registered with the Commerce Department reviews the lab and its models periodically, and the lab can pay that evaluator if it doesn't compromise its independence. Three kinds of risk are covered: bioweapons, AI-enabled cyberattacks, and a lab losing control of its models or agents. If the evaluator sees one coming, it has seventy-two hours to report to Commerce. Commerce then takes the evidence to a federal judge in D.C., who decides whether the release gets halted. And then there's the clause saying no other department, agency or officer of the United States can order a lab to stop releasing a model. Only Commerce can. That's a preemption clause, drafted by a company it would regulate. Anthropic says the bill reflects its Advanced AI Framework, which it's been pitching to federal and state lawmakers since June. To its credit, it names three specific harms instead of filing everything under AI safety. Anthropic is also doing something along these lines already. Ars Technica reported Sunday that METR, the evaluation nonprofit, now has a team embedded in Anthropic's office, with employee-like access to its hardware and internal systems. So it's demonstrating outside evaluation in its own office and asking Congress to make that the national template. Smart move. But if I'm building on these models, the question is who gets the stop button. Under this draft, that's one agency and one courtroom, and the lab gets to pay the referee.
Did we get something wrong today, or is there a story we should be chasing? Send it to aidailybriefing at lantern podcasts dot com.
That covers Muse Spark Open, Beam, Loughton and Anthropic's draft bill, and every source is in the show notes. AI Daily Briefing returns tomorrow. Thanks for spending part of your Monday with this Lantern Podcast.