Meta Releases Muse Spark Open Weights Under a Licence That Gates $20 Million-Revenue Resellers

Meta Platforms released the weights of Muse Spark Open on 15 October 2026, an 816-billion-parameter model with 42 billion active parameters per token and a 1-million-token context window, which the company calls "the most capable US open-weight model to date," according to The New Stack. The weights are free to download, but a licence requires companies with more than $20 million in annual gross revenue to sign a separate agreement with Meta before selling the model as a service.

That pairing of open weights and conditional terms is the real news. Meta is entering the contest against Chinese open-weight models with a system that leads on some agentic benchmarks and trails on coding, and it is doing so under a licence that stops short of open source.

What did Meta release?

Muse Spark Open is the open-weight sibling of Meta's Muse Spark family, which the company first launched in April as a break from its Llama line, AFP reported in a story carried by Khaleej Times. Meta describes it as built for "coding and agentic work," and says it "excels at long-horizon tasks, complex software engineering, and multi-step reasoning."

Distribution was broad from the first day. According to The New Stack, the model is available through the APIs of Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, and the weights can be downloaded from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft's Foundry Model Catalog. Mark Zuckerberg announced the release on X on 15 October, saying the weights were "releasing today."

How does the Muse Model License work?

Meta calls the terms the "Muse Model License" and says they carry "limited commercial terms, which allow many developers to use the model for free," according to The New Stack. Internal use and building the model into a company's own product are free.

Two conditions stand out. First, any company with more than $20 million in gross revenue over any consecutive twelve-month period 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 New Stack quoted the licence as saying. Second, products that use the model or a derivative must "prominently display 'Built with Muse Spark Open' in the user interface," including derivatives that were trained, fine-tuned or distilled from it.

The accompanying usage policy bars military and critical-infrastructure deployments, according to The New Stack, which noted that Meta added a similar carve-out for the U.S. government with Llama in 2024. The outlet also said the terms resemble the licence Moonshot AI uses for its Kimi K3 model.

How does it compare with Chinese open models?

Meta positions the model against Chinese systems that dominate open-weight use. AFP named Alibaba's Qwen, Moonshot AI's Kimi K3 and Z.ai's GLM models as the open-weight frontrunners.

On Meta's own published benchmarks, as described by The New Stack, Muse Spark Open beats the two U.S. open models it competes with, Thinking Machines' Inkling and Nvidia's Nemotron 3 Ultra, "often by 20 points or more." Against Chinese rivals the picture is mixed. The New Stack reported that it leads Kimi K3, GLM 5.3 and Qwen3.8 on agentic and tool-use tests including JobBench and MCP-Atlas and roughly ties them on SciCode. On coding it trails: on Terminal-Bench 2.1, Kimi K3 scores 85.0 and GLM 5.3 scores 83.9 against 76.4 for Spark Open, and on DeepSWE 1.1 the gap is 68 to 60.

These are vendor-reported comparisons. Independent evaluation is thinner and less flattering. The evaluation firm Vals.ai lists the model at 46.61% on its Vals Index, ranked 38th of 49 models. Its best result there is on the Finance Agent (v2) benchmark, where it scores 61.34% and ranks third of 79. It ranks 40th of 49 on Terminal-Bench 4.0 with 6.57%. Vals.ai also lists token pricing of $1.25 per million input tokens and $4.25 per million output tokens.

Why does the Apple connection matter?

Meta tied the launch to Apple hardware. Zuckerberg said the model "runs great on new Mac Studios" thanks to work with Apple and its chief executive, John Ternus, according to AFP. Ternus, who replaced Tim Cook in September, called it "amazing to run a model of Muse Spark Open's calibre and scale this efficiently," AFP reported.

The practical claim is narrower than the marketing. According to The New Stack, Meta says that two new Mac Studios with M5 Ultra chips and 512 GB of memory can be linked and run the model using Apple's MLX framework. The New Stack noted that Apple does not currently sell the 512 GB configuration to the public, though it expects it to be on sale in a few weeks, at a combined price it likened to a compact car. AFP likewise said the model still requires significant computing power.

Why does it matter for the open-model race?

Meta's strategy has changed in public. Llama made it the standard-bearer for open models, but AFP said that line lagged the industry, and the April Spark launch moved Meta toward closed frontier systems. On 12 October, three days before Spark Open, Meta unveiled two closed models, Wonder and Wonder Full, which it says rival the most advanced versions of ChatGPT and Claude, AFP reported.

What we don't know yet

Several things are unconfirmed. Meta's benchmark comparisons come from the company's own materials, and the sources reviewed here do not include a full independent replication on the same tests. Neither The New Stack nor AFP reported how many developers have downloaded the weights or how the licence terms will be enforced against resellers. And the full licence text should be read directly by any company near the $20 million revenue line, since the quotations above come from press coverage rather than the document itself.

What happens next?

Three things to watch. First, whether independent leaderboards move the model up from its current Vals Index position as more tests are run. Second, whether the 512 GB Mac Studio ships on the timeline The New Stack described, which determines how many teams can actually run the model locally. Third, whether cloud providers that already list the weights, including Amazon, Google and Microsoft, sign the separate agreement the licence requires for model-as-a-service offerings.