Meta released the downloadable weights for Muse Spark Open on Thursday, October 15, 2026 — an 816-billion-parameter model built for coding and long-running agent work that developers can run on their own hardware, including a pair of Apple Mac Studio desktops. Why it matters: it is Meta’s biggest open-weight release yet, and it puts a frontier-scale model outside the company’s own services for the first time in the Muse line.
October 15, 2026
What happened
Muse Spark has been a closed model since its debut in April. The new Open release changes that: the trained weights are now available to download from Hugging Face and through the model catalogs of Amazon Web Services, Google and Microsoft, and hosted access is offered by several inference providers, including Groq and Fireworks AI. Meta built this version from Muse Spark 1.2, which shipped in August, and added safety post-training intended to let the model operate without the guardrails of Meta’s own products around it. The company says the model cleared the release thresholds in its internal scaling framework, with a fuller evaluation report still to come.
The details that matter
The model is a mixture-of-experts system: 816 billion parameters in total, with about 42 billion active for any given token. It accepts text, images, audio and video, produces text, and supports a context window of up to 1 million tokens. Training drew on real developer workflows, including multi-file code changes and long debugging sessions, according to a technical write-up from Hugging Face.
Meta’s most eye-catching test ran the model locally, away from a data center. Two Mac Studio machines with Apple’s M5 Ultra chips, linked through Apple’s MLX framework with roughly a terabyte of memory between them, ran Muse Code while it fixed bugs and added features in an open-source Mastodon app, and reviewed video footage frame by frame to select shots for vertical video.
On benchmarks — all from Meta’s own table, so treat them as company-reported rather than independently verified — the model scored 61.8 on the JobBench agent test, ahead of the other U.S. open models Meta listed, and 89.2 on the MCP-Atlas tool-use test. On pure coding it trailed leading Chinese open models: its 76.4 on Terminal-Bench 2.1 sat below Kimi K3, GLM 5.3 and Qwen3.8 in the same table.
The license is not fully open
The weights ship under a new Muse Model License with limited commercial terms. Developers can adapt the model and ship it inside their own software without a fee, but products built on it must carry a Built with Muse Spark Open label, military and critical-infrastructure uses are excluded, and any company offering the model as a hosted service that has passed $20 million in gross revenue over 12 months needs a separate agreement with Meta. Meta’s previous open-weight release was the far smaller 30-billion-parameter Muse Glimmer in August, and its most capable Muse Wonder models, launched earlier this week, remain closed and API-only.
Why it matters
For readers, the practical takeaway is choice with a catch. A downloadable model of this size lets companies keep code and data on their own infrastructure and lets providers compete on price — but running it locally means very large-memory hardware, not a typical laptop, and the license terms need reading before any commercial launch. Everything AI, tested: we will be watching for independent benchmarks, because so far every headline number here comes from Meta itself.
Frequently asked questions
Is Muse Spark Open free to use?
The weights are free to download, and adapting the model or shipping it inside your own software carries no fee. Offering it as a hosted service above the revenue threshold in the license requires a separate agreement with Meta.
Can Muse Spark Open run on a normal computer?
Not realistically at full size. Meta demonstrated it across two high-memory Mac Studio desktops with about a terabyte of combined memory. Most users will reach it through a hosted provider instead.
How does it compare with Chinese open models?
In Meta’s own benchmark table, it led on agent and tool-use tests but trailed Kimi K3, GLM 5.3 and Qwen3.8 on the Terminal-Bench coding test. Independent results are not yet available.
Sources: Reuters, SiliconANGLE