Muse Spark Open License: Who Pays Meta and Who Doesn't

Meta put the weights of Muse Spark Open on Hugging Face on October 15, and the headline version of the story is easy: an 816-billion-parameter model, 42 billion active per token, a 1-million-token context window, built for coding and agent work. Meta calls it “the most capable US open-weight model to date.”

I spent more time on the license than on the specs, because that’s where “open” gets defined.

What the license lets you do

Most of it is generous. The license, officially the Muse Model License, says anyone may use, copy, modify and sell copies of the model, fine-tuning and distilling included, free of charge. Meta claims no ownership of what the model outputs. A startup that builds a coding tool on top of it, or a company that runs it internally, owes Meta nothing.

Two conditions apply to everybody. You must show “Built with Muse Spark Open” somewhere in the product: the interface, the docs or a settings menu. And anything trained or distilled using the model or its outputs counts as a “Derivative” that carries the same terms.

Who has to call Meta

The catch is Section 3. If you give third parties access to AI models through an API and they receive outputs, that’s “Model as a Service” in the license’s words. If you and your affiliates took in more than $20 million in gross revenue over any 12 months, you have to notify Meta and sign a separate written agreement before offering the model that way.

The number is total gross revenue, not revenue from the model, so a big company with a small inference sideline would trip it. The license says the clause doesn’t cover internal use, research, or using the model to power features of your own product that isn’t itself a model service.

The fence points at one business: the cloud and inference companies. Seven of them, including Groq, Fireworks AI and Together AI, are hosting it at launch, along with Baseten, Modal, Nebius and OpenRouter. How each of those deals is structured isn’t public, and I couldn’t find anything saying so.

The usage policy is long

Short version: read it.

Attached to the license is an acceptable-use policy. It bars military and warfare use, nuclear industries and weapons development. It also bars operating critical infrastructure, transportation technology or heavy machinery. It also says the model isn’t intended for anyone under 18.

Some of that is standard boilerplate. The infrastructure line deserves a second look from anyone planning industrial automation. I’m not a lawyer, and nobody here should treat my summary as legal advice. Read the actual text before you build.

The benchmark claim is Meta’s own

According to SiliconAngle, Meta’s own table has Muse Spark Open at 61.8 on JobBench, more than twice the 31.1 it lists for Inkling, the open model from Thinking Machines Lab. I haven’t seen an independent run. Meta said a full evaluation report is “available soon,” so the “most capable” label is a claim for now, the same way every launch-day leaderboard is.

It’s also a deliberate step below the top. According to The Next Web, Alexandr Wang said “because of both safety and capabilities, we will generally not open-source our most powerful models.” Muse Wonder, announced October 12, stays closed. So this is the best open model Meta is willing to hand over, not the best model Meta has.

Running it at home

Meta says two Apple M5 Ultra Mac Studios, each with up to 512 GB of unified memory, can hold the model. The Hugging Face model card puts the floor at 700 GB of GPU memory for the NVFP4 version and 2 TB for the full BF16 one. By those numbers, a pair of Studios fits the compressed checkpoint and not the full one. Meta’s post doesn’t say which version it ran, and the model card doesn’t either.

Where I land

The license is a better deal than I expected for small teams and a real negotiation for anyone who hosts models for a living. Calling it “open” is fair if you mean the weights are downloadable. It isn’t open in the way an Apache 2.0 model is, and Meta’s own August release, the 30-billion-parameter Glimmer, shipped under Apache 2.0, so the company knows the difference.

I’d guess the $20 million line is there to make sure Amazon, Google and Microsoft pay if they turn the model into a product. Those three list the weights in their catalogs now, though, and nothing public says what they’ve agreed to.

For the wider open-model race, see our post on Reflection AI’s Beam, the other US open-weight attempt announced this month.

Meta's Muse Spark Open is free to download. The license has a $20 million catch.