Meta Moves Its Open-Weight Threshold to Model Hosts

Meta released the weights of Muse Spark Open on October 15, describing it in its announcement as a model with “816B total parameters (42B active)” that can “run on your desktop using two Apple M5 Ultra Mac Studios.” The 42 billion is the number of parameters used for each token, not the model’s size in memory. The model card states the hardware plainly: “at least 700GB of VRAM for NVFP4 and 2TB for BF16.” The desktop in question is two machines that, Meta says, have up to 512GB of unified memory each.

The more consequential line is in the license. Section 3 of the Muse Model License concerns a licensee or its affiliates that “provide a third party with access to artificial intelligence models as a product or service, via API or a similar interface”—what the license calls Model as a Service. Charging is not part of that definition. If “the aggregate gross revenue of the Licensee and its affiliates exceeds twenty (20) million US dollars” in “any consecutive twelve (12) month period,” the licensee “must notify Meta and enter into a separate written agreement with Meta” “before offering the Software or any Derivatives as part of that Model as a Service business.” The revenue counted is the whole corporate group’s, not what the model earns. The section does not apply to “internal use, research, or any commercial activity that is not Model as a Service,” or to use of the model as a component of the licensee’s own product or service when that product is not itself a model service. A large company can build Muse Spark Open into its own app. To offer the model to others through an API, it first needs a contract with Meta.

Meta’s earlier open-weight licenses placed their extra commercial term elsewhere. The Llama 2 license of July 18, 2023 required a licensee to request a separate license from Meta if its products had more than 700 million monthly active users in the month before release. The Llama 4 license, effective April 5, 2025, kept the same 700 million threshold. Those licenses carried other limits as well—Llama 2, for one, barred using its outputs to improve other large language models—but their gate for a separate license was a user count high enough to reach only a handful of the world’s largest platforms. The new gate is an activity: offering the model to others through an interface, by any group of companies above a revenue level far below that.

Meta is a supplier in that market. On July 9, it announced that “For the first time, developers can begin building with Muse Spark 1.1 via the new Meta Model API, now in public preview.” Its developer site describes the API as “Direct, self-serve and affordable access to Muse models” and lists Muse Spark Open among them. In my reading, the license now places its extra gate in front of the kind of service Meta itself provides. That is a description of the structure; neither the announcement nor the license states why the line was drawn there.

The open-source world has seen a related design. Eight years ago to the day, MongoDB began releasing its database server under the Server Side Public License, applying it to versions “released on or after October 16, 2018.” The company said it had observed organizations, “especially the international cloud vendors,” begin “to test the boundaries” of its previous license; the new terms required anyone who made the software “available to third parties as a service” to publish the source code of the software used to run that service. MongoDB’s own FAQ acknowledges that software under the license “is not considered open source by the OSI.” The mechanisms differ: MongoDB attached a disclosure obligation, Meta a revenue threshold and a negotiated contract. The overlap is in the structure. Both licenses leave the software broadly usable and attach an extra condition to providing it to others as a service.

Meta’s strongest case is that its line is narrow. Internal use, research and features inside a company’s own products are exempt at any revenue, and groups below $20 million can serve the model without a separate agreement. The announcement names, on day one, seven “Hosted Inference Partners: Together AI, Fireworks, Baseten, Modal, Groq, Nebius, and OpenRouter,” and says the model “is available today through API platforms at Baseten.” These are the kind of businesses Section 3 addresses, and they are serving the model. Whether each crosses the threshold, whether any has signed a separate agreement and on what terms could not be confirmed from public documents; nor does either document explain how the $20 million figure was set. What the public can read is when the contract requirement begins, not what such a contract contains.

Other conditions surround the weights. On the Hugging Face page, access requires logging in and providing a “full legal name, date of birth, and full organization name with all corporate identifiers,” on a form that cannot be edited after submission. Products built on the model must prominently display the notice “Built with Muse Spark Open,” in the interface or in accompanying documentation, a website or a settings menu, much as Llama 4 required “Built with Llama.” The usage policy says the model “is not intended for individuals under the age of 18.” On safety, the model card summarizes Meta’s evaluations in four areas—content safety, agentic risk, privacy and preparedness—with results from Meta’s own testing, and warns that downstream modifications “could degrade safety.” The announcement says “a full evaluation report” will be “available soon.” Meta’s October 2 framework post concedes the general limit of any such testing for open weights: “Someone holding the weights can resample to get around a refusal, prefill outputs, or fine-tune refusal behavior away.”

None of this makes Muse Spark Open less useful to the researcher or the startup that downloads it; Meta’s framework calls open-weight models “a scarce and valuable resource for research.” What has changed since 2023 is where the license puts its extra gate. Llama’s stood in front of the very largest platforms. Muse Spark Open’s stands in front of large companies that would offer the model to others through an API, a market in which Meta is also a seller. A separate agreement is a negotiation, and whether Meta has used it to set terms, or to say no, is not public. For a model, “open” now describes who may run it freely. Who may serve it to others at scale is settled in contracts the public cannot read.