Meta releases Muse Spark Open, an 800B-parameter open-weight model built for Apple M5 Ultra Macs

Meta has released Muse Spark Open, an 800-billion-parameter open-weight model, with weights available now on Hugging Face. The model is optimized for Apple Silicon, but running it locally requires about 1TB of memory.

Key specifications

• Parameters: 800 billion
• Release date: October 15, 2026
• Distribution: Open weights via Hugging Face
• Memory requirement: ~1TB for local inference
• Context window: not disclosed in available reporting
• API pricing: not applicable to the weights download; hosted pricing not yet disclosed
• Training cutoff: not disclosed

Hardware requirements

Meta says Muse Spark Open is the first model optimized to run locally on Apple's M5 Ultra Mac Studio. According to Meta, the M5 Ultra's integrated GPU can address up to 512GB of unified memory per machine, so two Mac Studios can hold the full model. Meta says the pair behaves "like one supercomputer with a terabyte of unified memory."

Two configurations are possible:

• Two 512GB M5 Ultra Mac Studios. Apple has not announced pricing or a specific ship date for the 512GB configuration. It says only that the machine will arrive in "late October."
• Four 256GB M5 Ultra Mac Studios. Engadget estimates this setup at roughly $38,000.

Mark Zuckerberg said the model was tuned in collaboration with Apple.

Performance claims

Meta calls Muse Spark Open "the most capable US open-weight model to date." The company says its benchmarks show the model outperforming two leading Chinese open models, GLM 5.3 and Kimi K3, on agentic tasks and visual reasoning. Meta has not published numerical scores in the material available so far, and the results have not been independently verified. The visual reasoning claim suggests the model accepts image input.

The release follows Meta's stated position that the US should not let Chinese labs lead in open-source AI. It also expands Meta's open-weight lineup by opening up the Muse Spark line.

Not yet disclosed

• License terms
• Context window
• Specific benchmark names and scores
• Training data cutoff
• Whether the 800B figure reflects total or active parameters

What this means

The 1TB requirement puts Muse Spark Open out of reach for individual developers. It is aimed at organizations that can spend tens of thousands of dollars on hardware to avoid recurring token costs and keep data on-premises. Any such calculation depends on Meta's benchmark claims holding up, and independent evaluations will determine that.

The release is also a strategic signal. Tying a frontier-scale open model to Apple's unified memory architecture makes high-end Macs a credible local inference platform for models at this scale. Cloud and datacenter GPU deployments still dominate large-model serving, though, and Apple's missing 512GB Mac Studio pricing leaves the real cost of the two-machine setup unknown. Expect third-party quantized versions to appear on Hugging Face. They would lower the memory barrier at some cost in quality.