Meta releases Muse Spark Open weights: an 816B open-weight model built to run on Mac Studios

Meta released the weights for Muse Spark Open on 15 October 2026 — 816 billion total parameters with 42 billion active per token, a 1-million-token context window, natively multimodal — calling it the most capable US open-weight model to date, and optimised to run on two Mac Studios with M5 Ultra chips.

15 October 2026 · News · Verified against the original sources on 3 November 2026

What happened

  • Meta unveiled Muse Spark Open on 15 October 2026, releasing its weights and describing it as the most capable US open-weight model to date.
  • Muse Spark Open is a natively multimodal model with 816B total parameters (42B active) and a 1M-token context window, built for coding and agentic work, long-horizon tasks, complex software engineering and multi-step reasoning, according to Meta’s announcement as quoted in launch reporting.
  • Meta says the model is optimised for Mac Studio with M5 Ultra using Apple’s MLX framework: each Mac Studio offers up to 512 GB of unified memory, so two machines hold the full model and behave like one supercomputer with a terabyte of unified memory.
  • Meta reports testing agentic workflows running entirely on-device with Muse Code powered by Muse Spark Open, including shipping iOS changes to the open-source Mastodon client Ice Cubes and driving DaVinci Resolve end to end.
  • The weights are available through Hugging Face. The release uses Meta’s own Muse Model License, not an unrestricted open-source licence: launch reporting notes attribution conditions, including a “Built with Muse Spark Open” notice, and a separate agreement for large model-as-a-service businesses. Read the licence itself before building a product on the weights.

What it means

  • For most AiTopMix readers, Muse Spark Open is not a tool you open in a browser — it is weights to download and run, or call through a hosted provider. Its practical importance is that a frontier-class open-weight model now exists for coding agents and long-horizon work without a per-token API bill, if you have the hardware or a hosted route.
  • The Mac Studio angle is the surprise: Meta and Apple worked together on MLX support so a desktop pair can run a model of this size. The 512 GB Mac Studio configuration was not yet on sale at announcement, according to launch reporting, so the desktop setup was a promise more than a purchase on day one.
  • Benchmark claims in Meta’s announcement are vendor-reported. Independent evaluations reported at launch placed the model behind rival Chinese open models on some coding benchmarks while ahead on tool-use tests. Match the model to the workload; do not take any single score as a verdict.

What we don’t know yet

  • Meta’s full safety evaluation report was described as coming soon at launch. The exact commercial limits of the Muse Model License need a direct read of the licence text.

How we covered this: facts above come only from the sources listed below. Vendor benchmark figures are the vendor’s own, labelled as such. AiTopMix did not independently test the products in this story.

Sources

Checked on 3 November 2026. If a source page has changed since, the source is the final word.

For coding tools in the directory, see Cursor vs GitHub Copilot vs Windsurf vs Replit.

← All news Back to the directory