Meta has an open-weight AI model back near the top of a public 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.

The “US” qualifier, however, is doing considerable work in that claim. The model leads among American open releases, while global open-model rankings remain a more crowded and complicated story.

What Meta Shipped

Meta Superintelligence Labs introduced Muse Spark Open on October 15, 2026, billing it as the most capable US open-weight model released so far.

The headline specification 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 (MoE) design: rather than engaging the full network for every request, a routing mechanism sends each input to a small set of specialized parameters, much as a hospital's triage system directs each patient to the few relevant specialists instead of the entire staff.

The context window spans one million tokens, enough room to load a sizable codebase or a long stack of documents in a single session. Muse Spark Open is natively multimodal on the input side, accepting 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 use. Those three focus areas are the same categories its published benchmarks and the Code Arena board measure, which gives readers a direct way to weigh the “most capable” billing against the numbers.

Availability and Hardware

The weights are published on Hugging Face in BF16 and NVFP4 formats.

One of the more striking claims concerns 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, a setup that carries roughly one terabyte of unified memory. For organizations that need workloads kept in-house, that is a notably compact footprint for a model with 816 billion total parameters.

For teams that prefer the cloud, distribution is broad. The model is available through Baseten, Fireworks AI, Google Cloud Model Garden, and Amazon SageMaker, giving teams several established routes to run the same weights.

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 that stands out. Arena's WebDev board ranks models on building web applications, a practical test of coding ability. Leaderboard placements are snapshots as well — positions shift as new models are evaluated — so the figure describes the field as it stands today rather than a fixed rank.

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. That score is the concrete data point behind the model's stated tool-use emphasis.

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. That mixed record is the context for the headline claim's scope: the WebDev lead holds among US open releases, while head-to-head results against the broader international field vary by benchmark.

How the Release Came About

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. From here, the items to watch are measurable: movement on the Code Arena board as additional models are scored, and third-party results from the published weights, which can be checked directly against the figures in Meta's announcement.