Meta has released Muse Spark Open, an 816-billion-parameter mixture-of-experts model with 42 billion parameters active per token, a 1-million-token context window and native multimodal input. The weights are downloadable under the Muse Model License, which permits limited commercial use, and Meta says the model was trained and developed in the United States. It is the company's strongest open-weights release to date and, by several independent measurements, the strongest open model yet for regulated professional work.

The release landed with unusually broad day-one availability. Fireworks is serving Muse Spark Open on US-only serverless capacity with Zero Data Retention by default and no US price premium, with reserved throughput for production and serverless LoRA fine-tuning on the same platform. Baseten, Groq, Modal, Nebius, OpenRouter and Together also carry the model, and Meta says it runs locally on as little as two Apple M5 Ultra Mac Studios through MLX.

Where it scores

The headline results come from independent evaluators rather than Meta's own harness. On Vals AI's leaderboard for Harvey's Legal Agent Benchmark, which tests completing legal work with files and tools, Muse Spark Open ranks first of 75 models at 25.4%, nearly four times the 6.7% pass rate of Claude Opus 5. On Rogo's Big Finance Bench, a set of 928 real financial workflow questions, it scored 53.3%, the top open-weights result on Rogo's board and ahead of Claude Opus 5 at 49.2%. Broken down by difficulty in Fireworks' Specialized Intelligence Index runs, it beat Claude Opus 5.5 at every level and edged out GPT-6 Astra on the hardest tasks, 17.3% to 16.5%.

"At Rogo we run every new model through Big Finance Bench," said Rogo's Alex Wang, who called the emergence of high-performing American open-weights models part of a broader shift in who can build on frontier-class systems. Factory's Zimu, a member of technical staff at the coding-agent company, pointed to "strong agentic engineering capability," saying the model can navigate repositories, implement nontrivial changes and validate complex coding work.

Two more applied results stand out for cost-sensitive deployments. On Vals AI's Finance Agent v2, Muse Spark Open scored 60.6%, within a point of Gemini 3.8 Flash at roughly 60% lower cost per test. On the Vals MedScribe clinical documentation benchmark it reached 90.1%, within a point of Claude Opus 5's 91.0%, at 75-83% lower token prices. Artificial Analysis' AA-Omniscience measurement put its hallucination rate at 33%, below Claude Opus 5.5's 59% and GPT-6 Astra's 51%. Vals AI and Artificial Analysis measured Meta's API release of the model.

On standard public benchmarks, Fireworks' validation runs recorded 91.1% on AIME26, 89.7% on GPQA Diamond and 84.3% on SWE-bench Verified, with image understanding above 80% on OCRBench and MMMU Pro. Meta's own Terminal-Bench 2.1 result is 82.9%, consistent with the long-horizon coding and agent positioning.

Why it matters

Open-weights models have spent two years trailing the frontier on exactly the tasks enterprises pay the most to get right: legal review, financial analysis and clinical documentation, all under data-handling constraints. A US-trained open model that leads a legal agent benchmark outright, pairs with Zero Data Retention serving, and fits on two desktops changes the procurement math for regulated industries that could not send data to a closed API and previously had no open option near the top of the leaderboard.

The caveats are real. The Muse Model License is not a permissive open-source license, and its commercial terms will exclude some uses. Benchmark rankings from vendor-run harnesses deserve independent replication, and a 25.4% top score on the legal benchmark also means three in four tasks still fail. But as of today, teams evaluating open weights for regulated work have a new default starting point.