Meta releases 816B-parameter Muse Spark Open for coding and agent work
For IT teams, the more important architectural question is whether Muse Spark Open’s strengths align with real workflows: long-context software maintenance, multi-step tool invocation and multimodal input. Those capabilities are useful only if organizations can connect the model cleanly to source control, CI/CD, issue trackers, artifact stores and governed MCP-compatible tools. Tool-use benchmark wins do not remove the need for permission scoping, auditability, rate limiting and rollback paths when an agent is allowed to act across multiple systems.
The license terms also make deployment strategy a first-order concern. Teams embedding the model into internal products may evaluate it differently from vendors planning hosted AI services, where commercial thresholds and branding requirements could affect packaging and margin. In practice, that pushes buyers to compare three paths: self-hosted open weights for control, inference providers for speed, or closed API models for lower operational burden.
Before adopting, practitioners should validate four things in their own environment: sustained long-context performance, tool-call reliability under failure conditions, total infrastructure cost at target concurrency, and how much additional safety and policy enforcement is needed outside Meta’s own guardrails.
Meta Platforms Inc. today released open weights for its Muse Spark artificial intelligence model, which had been proprietary since its April debut.
The release, called Muse Spark Open, is built for coding and long-running agent work. Meta describes it as the most capable open-weight model to come out of a U.S. lab so far. Only 42 billion of the model’s 816 billion parameters switch on for any given token. Prompts can mix text with images, audio or video, and Meta has set the context window at up to 1 million tokens.
The starting point was Muse Spark 1.2, shipped in August, and Meta then added safety post-training so the model holds up without the guardrails of Meta’s own services. Meta said Muse Spark Open cleared the release thresholds in its Meta Superintelligence Scaling Framework, with a full evaluation report still to come. A technical writeup from Hugging Face Inc. says training leaned on what Meta calls “real developer workflows,” including multi-file refactors and long debugging sessions.
Meta also tested the model away from the data center, on a pair of Apple Inc. Mac Studio desktops with M5 Ultra chips that Apple’s MLX framework linked into one system. The two machines had about a terabyte of memory between them. Muse Code ran on that setup while the coding terminal worked through bug fixes and new features for the open-source Mastodon client Ice Cubes. Footage in DaVinci Resolve was the other test, and the model went through it frame by frame to pick out and crop the best shots for vertical video.
In Meta’s own benchmark table, Muse Spark Open scored 61.8 on the JobBench agent test, more than twice the 31.1 recorded by Inkling, the U.S. open model from Thinking Machines Lab Inc. Nvidia Corp.’s Nemotron 3 Ultra trailed further at 27.2.
Chinese open models are stronger at coding. Kimi K3 led the Terminal-Bench 2.1 coding test at 85, and GLM 5.3 and Qwen3.8 both finished above Muse Spark Open’s 76.4.
On tool use the order flips, and Muse Spark Open’s 89.2 on MCP-Atlas beat every rival Meta listed, the Chinese models included. The benchmark gives a model plain-language tasks without naming the tools needed to complete them. Working out which of 220 tools across 36 live Model Context Protocol servers to call, and in what order, is left to the model.
The weights come with a new Muse Model License carrying what Meta calls “limited commercial terms,” and no fee applies to developers who adapt the model or ship it inside their own software. Meta requires a separate agreement from any company that sells the model as a hosted service and has passed $20 million in gross revenue over a 12-month period. Products built on it have to show a “Built with Muse Spark Open” label. Military and critical infrastructure uses are off-limits.
Meta’s previous open-weight release was the much smaller Muse Glimmer, a 30 billion-parameter model that went out in August. The company’s most capable systems remain closed, and the Muse Wonder models introduced Monday are available only through the Meta Model API and Muse Code.
Seven inference providers offer hosted access at launch, Groq Inc. and Fireworks AI Inc. among them. The weights can also be downloaded from Hugging Face or from the model catalogs of Amazon Web Services Inc., Google LLC and Microsoft Corp.
Meta releases 816B-parameter Muse Spark Open for coding and agent work
Enjoyed this article? Sign up for our newsletter to receive regular insights and stay connected.