Meta has open-sourced the weights for Muse Spark Open and worked with Apple to make the 816-billion-parameter model run locally on Mac Studio hardware.
Muse Spark Open weights are now available on Hugging Face
The model activates 42 billion parameters per inference, supports a 1 million-token context window, and is geared toward coding and complex agent tasks. Its weights were listed on Hugging Face on Oct. 15, and developers can apply for access after signing in and accepting the access conditions.
Meta and Apple used MLX to spread the model across multiple machines
Meta and Apple have previously clashed several times over privacy policy. In this case, Mark Zuckerberg publicly thanked Apple CEO John Ternus, who also responded publicly. The two companies used Apple’s MLX framework to distribute the model across multiple Mac Studio systems.
A single Mac Studio cannot hold the full deployment
Even when compressed to 4-bit precision, the model still needs about 493GB of memory. After accounting for inference overhead, one Mac Studio is not enough. A practical deployment requires at least two Mac Studio units equipped with M5 Ultra chips and 512GB of unified memory each, bringing hardware costs to more than $40,000.
Apple shared performance figures and Meta set license terms
According to Apple’s test results, four M5 Ultra Mac Studio machines running together generated about 37 tokens per second with an 8,192-token context.
Meta is using a custom Muse Model License that allows developers to modify, commercialize, and distribute the model, but it requires products or related documentation to state that they use Muse Spark Open. If a company and its affiliates post more than $20 million in total revenue over a rolling 12-month period and want to offer the model externally as an API service, they must sign a separate agreement with Meta first.