Meta has an open-weight model back near the top of a leaderboard. Muse Spark Open (Max) scored 1511 points in Arena’s Code Arena: WebDev rankings, making it the highest-ranked US open model on that board.
That qualifier, “US,” is doing a lot of work. The model leads among American open releases. Global open-model rankings are a more crowded and more complicated story.
What Meta actually shipped
Meta Superintelligence Labs introduced Muse Spark Open on October 15, 2026. The lab is billing it as the most capable US open-weight model released so far.
The headline spec is size. The model has 816 billion total parameters, but only 42 billion of them activate for any given token.
That gap comes from its Mixture-of-Experts design, or MoE. Think of it as a large hospital staffed with specialists. Every patient doesn’t see every doctor; a triage system routes each case to the few experts who actually matter.
The context window is one million tokens. That is enough room to feed in a sizable codebase or a long stack of documents in a single session.
Muse Spark Open is natively multimodal on the input side. It accepts text, images, video and audio, though its outputs are primarily text.
Meta says the model was tuned for complex software engineering, agentic workflows that run over long horizons, and tool use.
The weights are published on Hugging Face in BF16 and NVFP4 formats.
The Mac Studio flex
One of the more striking claims is where the model can run. According to the release details, Muse Spark Open works on two paired Apple M5 Ultra Mac Studios using Apple’s MLX framework.
That setup carries roughly one terabyte of unified memory.
For teams that prefer the cloud, distribution is broad. The model is available through Baseten, Fireworks AI, Google Cloud Model Garden and Amazon SageMaker.
Meta also says the open release received additional safety post-training compared with earlier variants.
How it stacks up
The 1511 score in Code Arena: WebDev is the number Meta’s fans will quote. Arena’s WebDev board ranks models on building web applications, a practical test of coding ability.
On tool use, Meta’s own figures are strong. Muse Spark Open scored 89.2 on MCP-Atlas, a benchmark focused on how well models work with external tools.
The coding picture is less tidy. Meta’s own benchmark results show mixed performance against international competitors, with the model trailing on certain coding evaluations.
How we got here
Muse Spark did not start life as an open model. The original closed Muse Spark appeared in April 2026, followed by Muse Spark 1.1 in July.
The open release fulfills a commitment to make the Muse family available in open form.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.