Meta's Open Model Isn't Open. That Is the Point.

Washington spent this week arguing about who gets to supervise AI. President Trump's Super Intelligence Task Force held its inaugural meeting on October 15, setting up committees and commissioning a risk study. The same day, Reuters reported that Anthropic has been quietly pitching Capitol Hill a draft safety bill it wrote itself, one that would let the government halt a model only if it poses a foreseeable risk of more than $1 billion in damages or of killing or seriously injuring more than 50 people. And on Wednesday, eight state attorneys general filed parallel lawsuits accusing OpenAI of marketing ChatGPT as safe for kids while it harmed vulnerable users.

All of that is a debate about companies, APIs, and press releases. Meanwhile, the most consequential AI event of the week was something none of those proceedings can touch. On October 15, Meta published the weights of an 816-billion-parameter AI model, free for anyone to download. You cannot subpoena a download. You cannot put a kill switch on a file that is already on ten thousand hard drives.

Once the weights are out, they are out

Muse Spark Open is natively multimodal, tuned for coding and agentic work, with a 1-million-token context window, 816 billion total parameters, and 42 billion active per token (TechCrunch, MacRumors, October 15). The weights are downloadable from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Seven hosted platforms, Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, offer it as an API from day one. Meta says it ran the model through its own Superintelligence Scaling Framework before release and promised a full evaluation report soon. But the fundamental act, publishing the weights, is irreversible.

This is what diffusion actually looks like. Every conversation in Washington this week treated AI as something that happens inside a vendor's cloud, where a regulator or a judge can reach it. Open-weight releases move the frontier onto hardware the vendor does not own and cannot monitor. The policy world is designing locks for doors that the industry is quietly removing.

A kill switch works on a company. It does nothing to a file.

"Open" is doing a lot of work

Meta calls this an open release. It is not open source, and the distinction matters. Muse Spark Open ships under Meta's own Muse Model License, which carries, in Meta's own words, limited commercial terms. Companies that sell AI models as a service and take in more than $20 million in annual revenue must notify Meta and sign a separate agreement. Every product built on the model must display a "Built with Muse Spark Open" notice (AIStockWire, October 2026).

That is open weight, not open source: the weights are downloadable, the freedoms are conditional. Meta gets the halo of openness, a tax on its competitors' API businesses, and a developer base building on its architecture, while keeping legal leverage over the biggest commercial users. It is a brilliant strategy. It is also, definitionally, not open.

Then there is the headline hardware story. Meta engineered the model to run on a pair of Apple M5 Ultra Mac Studios, with Apple's MLX framework splitting the model across two machines that, in Meta's words, behave "like one supercomputer." Apple says four Mac Studios generate more than 36 tokens per second. It is a genuine technical achievement, and it is also the most expensive definition of "on your own hardware" I have seen this year. The 256GB Mac Studio already costs over $10,000, and the 512GB configuration the two-machine setup needs ships later in October at a higher price (MacRumors). The cheapest honest price for "running it yourself" is somewhere past $20,000 in Apple hardware. Meta commoditizes the model. Apple monetizes the box.

The race is not where you think it is

Meta calls Muse Spark Open "the most capable US open-weight model to date," and that phrasing deserves a second look. Per Artificial Analysis scores reported this week, the model sits at 39 on the Intelligence Index, the highest of any open-weight model built outside China. Seven Chinese models score higher, led by Xiaomi's MiMo-V2.6-Pro at 46, with DeepSeek V4.1 Flash tying Muse Spark Open at 39. For reference, Meta's own closed model, Muse Wonder, scores around 56.

Every one of these numbers is vendor-reported or partner-scored, not independently verified, and Meta's own comparison chart conspicuously omits DeepSeek V4. Treat the exact rankings with caution. But the shape of the picture is consistent: in the open-weight category, which is the category where capability actually spreads beyond anyone's control, the United States is playing catch-up to China. The AI race Washington talks about is measured in data center announcements and frontier API benchmarks. The race that is actually happening, in weight files and Hugging Face downloads, is one America is currently losing.

That reframes everything else. If open weights are the real distribution mechanism of AI power, then the lever that matters is not which company signs an accord or which task force writes a report. It is who publishes the best weights, under what license, and on whose hardware they run. Meta's release is a master class in all three: flood the zone with capable weights, keep the license terms, and aim the whole thing at Apple's silicon.

The honest version of this week's AI conversation would start there. Stop asking whether the labs can be trusted to grade their own homework. Start asking what it means that the homework is now downloadable, licensable, and $20,000 in Mac Studios away from running in your office, while the best versions of it are coming from Shenzhen and Hangzhou. The future of AI is not being decided in committee rooms. It is being decided in model cards and license files, and this week, the committees were not even in the room.