Meta has released the weights for Muse Spark Open, a natively multimodal AI model with 816 billion total parameters. “Today, we’re releasing the weights for Muse Spark Open, the most capable US open-weight model to date,” the company said in its announcement.
The standout detail is where it runs. Meta optimized the model for Apple’s Mac Studio with the M5 Ultra chip, using Apple’s MLX framework. Two linked Mac Studios, each with up to 512GB of unified memory, hold the entire model.
Meta CEO Mark Zuckerberg announced the release on X, writing that it “runs great on new Mac Studios thanks to work with @johnternus and Apple.”
Muse Spark Open — weights releasing today — runs great on new Mac Studios thanks to work with @johnternus and Apple. pic.twitter.com/U2wJNrKBOg
— Mark Zuckerberg (@finkd) October 15, 2026
Apple CEO John Ternus replied: “Congrats to the team! Amazing to run a model of Muse Spark Open’s caliber and scale this efficiently, right from your desk. Mac Studio keeps pushing the boundaries of on-device AI!”
Built for coding and agentic work
Only 42 billion of the 816 billion parameters are active per token, using a mixture-of-experts design. The model takes text, image, video and audio as input and returns text, with a one-million-token context window. Meta says it excels at long-horizon tasks, complex software engineering and multi-step reasoning.
To show it off, Meta paired the model with Muse Code, its AI coding terminal, and shipped real iOS changes to Ice Cubes, an open-source Mastodon client. The work covered bug fixes, new features and a full build.
It also drove DaVinci Resolve end to end, reviewing footage frame by frame to find the strongest moments and crop them vertically.
Desktop AI with a steep hardware bill
Running it locally is not cheap. The model needs at least 1TB of RAM, which means two maxed-out Mac Studios working as one machine. Apple has not announced pricing for the 512GB Mac Studio, due in late October, but four 256GB M5 Ultra units would cost roughly $38,000.
Meta AI chief Alexandr Wang told Axios that the Mac Studio’s memory makes a model this size practical to run, unlike on most personal computers. Because processing happens on the device, Meta has no access to user prompts or usage data.
The weights are available in BF16 and NVFP4 on Hugging Face, plus Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. API access runs through providers including Fireworks AI, Groq and Together AI.
Open weights with strings attached
The model ships under the Muse Model License. Many developers can use it free, but companies selling model access through an API with more than $20 million in gross revenue over 12 months need a separate agreement with Meta.
The open version also went through extra safety post-training and met the release thresholds of Meta’s Superintelligence Scaling Framework, which governs how Meta releases powerful models.
On benchmarks, Meta’s own scorecard has Muse Spark Open leading 6 of 11 tests against rival open-weight models, including Chinese models GLM 5.3 and Kimi K3 on agentic tasks and visual reasoning. But on Terminal-bench 2.1, a software engineering test, it scored 76.4 and trailed Kimi K3 at 85.0.
Why Meta is doing this
The release continues Meta’s open-source push. Zuckerberg has argued that advanced AI should not be concentrated among a small number of companies, and Meta has said the US should not let Chinese labs lead on open-source AI.
Muse Spark Open is derived from the Muse Spark model released in August, with added safety training for standalone use.
