Fireworks adds Meta’s Muse Spark Open with US-only hosting and no location premium

Developers can now call Meta’s Muse Spark Open through Fireworks without running the model themselves. Fireworks announced the availability on October 20, 2026, offering US-only server capacity without an added US premium. Its pitch combines a large context window for extended work with benchmark results aimed at legal, financial and clinical users.

A large model, accessed through a hosted service

Muse Spark Open is an open-weight model from Meta. Fireworks describes it as a mixture-of-experts system with 816 billion parameters in total, but 42 billion active for each token it processes. That distinction separates the model’s overall size from the portion used at each step; it does not activate all 816 billion parameters for every token.

Its context window holds up to one million tokens—the material it can work with in a request. Fireworks positions that capacity for long coding sessions and multi-step agent work that must remain coherent over time, rather than only short exchanges. The context limit describes capacity; the sustained-work capability is the company’s stated goal.

The serverless API is the starting point. Fireworks also offers reserved throughput for production workloads and serverless LoRA training for customization. It says zero data retention is enabled by default. The announcement directs developers to the model page for current prices, but gives neither specific API rates nor the model’s license terms.

Professional-task rankings need their denominators

Fireworks makes its strongest competitive case around regulated professional work. The legal test it highlights involves completing work with files and tools. Its finance comparison includes results from Fireworks’ Specialized Intelligence Index, where the company says models use the same evaluation harness—the surrounding software that runs the task.

• Legal: Fireworks reports first place among 75 models on Vals AI’s Harvey Legal Agent Benchmark, with a 25.4% pass rate versus 6.7% for Claude Opus 5. First place still meant passing roughly a quarter of the benchmark’s tasks.
• Finance: In Fireworks’ Big Finance Bench runs, Muse Spark Open beat Claude Opus 5.5 at every difficulty level. On the hardest tasks, it scored 17.3%, compared with 16.5% for GPT-6 Astra and 6.6% for Opus 5.5.

Other reported results cover financial agents and clinical documentation. Fireworks cites 60.6% on Vals AI’s Finance Agent v2, within a point of Gemini 3.8 Flash at about 60% lower cost per test. On Vals MedScribe, it reports 90.1%, versus 91.0% for Claude Opus 5, with token prices 75–83% lower.

Checking the model on Fireworks’ own stack

Fireworks ran separate launch checks using the final open weights on its production serving system. It reports 91.1% on AIME26, 89.7% on GPQA Diamond and 84.3% on SWE-bench Verified. Image-understanding scores exceeded 80% on OCRBench and MMMU Pro. It also says the same weights produced matching scores on SGLang, an open-source inference engine.

Those checks should not be merged into one coding leaderboard. Meta’s separately reported Terminal-Bench 2.1 score is 82.9%, but each model in that comparison used its own agent harness. Fireworks explicitly warns that agent results depend on the harness and that its figures are not directly comparable with Meta’s.

A separate result addresses guessing rather than coding. Fireworks cites Artificial Analysis’ AA-Omniscience evaluation: a 33% hallucination rate, compared with 59% for Claude Opus 5.5 and 51% for GPT-6 Astra. It attributes the lower rate to declining uncertain questions instead of guessing—a claim about when the model answers, not simply how many answers it gets right.