Meta Muse Spark Open, an 816-billion-parameter model, went public on Thursday, October 15, and anyone can download its weights. Meta says it is the "most capable" open-weight model in the United States, and it is built to run locally on Apple silicon Macs. Those are Meta's claims, relayed by TechCrunch, and the performance numbers come from evaluations Meta released itself.

The timing is the second part of the story. Meta launched its closed model, Muse Wonder, on October 12, then released the open companion three days later. Bloomberg reports that Meta kept its most advanced technology closed and that Spark Open is less advanced than Muse Wonder.

What Meta released in Muse Spark Open

Muse Spark Open is a mixture-of-experts model. AI Weekly reports 42 billion active parameters inside the 816 billion total, which means only a fraction of the network works on any given token. Meta also cites a 1-million-token context window, and its Superintelligence Lab released the model.

Meta describes the weights as the settings that let developers run and adapt the model. That is what separates this release from a hosted product. Open weights let anyone download, modify and run a model on their own hardware. A service like ChatGPT or Claude works the other way: you send a request to the provider's servers and pay for the output.

Bloomberg says the release fulfills an earlier Meta commitment to make its technology widely available. Read alongside the closed Muse Wonder launch, the pattern is clear. Meta is putting a capable model in developers' hands while keeping its strongest work behind closed doors, which Bloomberg frames as a way to turn AI investment into revenue.

Which Macs can run Meta's open model

Meta built the model to run on Apple silicon Macs, and TechCrunch reports that Meta and Apple worked together on the launch. The Neuron says the model can run on two Mac Studios. AI Weekly describes those machines as paired M5 Ultra Mac Studios.

Those hardware details come from secondary coverage, and the sources do not say whether the two-machine setup applies to the full model or to a compressed version. Even a compressed copy of an 816-billion-parameter model needs a large pool of memory, so this is not a laptop project for most people. Mac owners with high-memory machines are the most direct audience. Everyone else can still read the license and model card, and follow hands-on tests as they appear.

The sources reviewed for this story include no independent hands-on test, and the benchmark rankings are Meta's own. Until someone runs the model on a real machine and publishes the results, the hardware claim remains a company statement relayed by the press.

Privacy is where local deployment could matter most. Enterprises weighing data privacy may prefer a model whose weights run on their own machines, since no data has to leave for a cloud provider. That advantage only holds if the hardware is affordable enough to justify the switch.

What the Muse Model License allows

AI Weekly reports the model ships under the Muse Model License, which permits free use for many developers. The qualifier matters. The sources do not spell out commercial-use limits, and they do not say whether very large users face restrictions.

Open weights and open source are not the same thing. A license can permit downloads and local use while limiting commercial deployment, redistribution or use by very large companies. A startup building a paid product on this model needs the full license text, and possibly a lawyer's review, rather than a summary from a news roundup.

Meta's broader terms are a separate issue. USA 24 has reported that Meta's October 30 terms may let it scrape and train on news sites. That story concerns Meta's platforms rather than this model, but anyone evaluating Meta's AI products should read both sets of terms.

How Muse Spark Open compares with Muse Wonder

Muse Wonder launched October 12 as Meta's first frontier model. Local AI Zone reports API pricing of $4 per million input tokens and $20 per million output tokens. Those figures come from an aggregator, so check them against Meta's own pricing page before budgeting.

The open release shifts the cost calculation. Hosted APIs charge per token, while self-hosting trades those fees for hardware, power and engineering time. A team that already owns high-memory Macs may find local deployment cheaper over time. A team without that hardware will probably find a paid API simpler.

The release also arrived during a busy news cycle. BenchLM's live tracker lists Muse Spark Open among five confirmed model releases in the prior 24 hours, alongside Nvidia's Nemotron 3.5 Super VL. Each new open model gives developers another option to weigh against paid services.

Policy is watching too. China has proposed export controls on AI model weights and source code, a move that could shape how open releases cross borders. USA 24's coverage lays out the proposal.

Policy backdrop: a White House task force on AI risk

CryptoBriefing reports that the Super Intelligence Force, a White House task force, held its first meeting on October 15 and began a 120-day review of AI cybersecurity, biosecurity and model control risks. That review is separate from the Meta release. It still matters here, because downloadable weights are very hard to recall once they are public. USA 24 has also examined who bears liability when AI agents misbehave, including a warning the task force issued to developers.

Practical steps before you download Muse Spark Open

  • Read the Muse Model License in full before any commercial use, and get legal review if you plan to ship a product.
  • Open Meta's model card and check each benchmark claim. The performance numbers are Meta's own, and the sources reviewed did not include Meta's primary announcement, so confirm details against Meta's posts.
  • Verify hardware requirements before installing, including whether the version you download is quantized and how much memory it needs.
  • Start with non-sensitive test data, and record the hardware and settings you used so your results can be compared with others.

What to watch next

Three things will settle most of the open questions. Independent benchmarks will show whether the performance claims hold against other open-weight models. The full license text will show how far commercial use stretches. Hands-on tests on Mac hardware will show what it takes to run the model in practice. Watch as well whether Meta releases smaller versions, and how the White House review treats open weights as its 120-day clock runs.