Apple says Muse Spark Open does not fit on a single Mac Studio in its tested setup. Meta announced the model on October 15, 2026, and Apple reported inference measurements using two- and four-machine configurations.
Meta CEO Mark Zuckerberg’s post credited Apple and John Ternus for the work on Mac Studio support.
What Muse Spark Open is
Muse Spark Open is Meta’s open-weight, natively multimodal model, designed for coding and agentic tasks—work in which software carries out actions toward a goal. Meta lists 816 billion total parameters, 42 billion active parameters, and a context window of 1 million tokens. It says the model was optimized to run on Apple silicon through MLX, Apple’s machine-learning framework. (Meta’s announcement)
Why Apple’s setup uses multiple Mac Studios
The model’s memory demands explain the multi-Mac approach. Apple says its BF16 model parameters alone require more than 1.6 TB of RAM. In NVFP4, a compressed numerical format used for the evaluation, the model footprint is about 493 GB before the additional memory needed for intermediate inference computations. Apple says the model therefore does not fit on one Mac Studio in its tested setup. (Apple’s technical report)
That distinction matters: a model’s active-parameter count is not the amount of memory needed to load and run it. Meta describes a two-Mac Studio route; Apple also measured a four-machine configuration.
What Apple measured
Apple measured decoding after an 8,192-token context, generating 128 tokens. The results below are Apple’s measurements for those configurations, not a general speed guarantee for other workloads or setups. (Apple’s technical report)
| Mac Studio configuration | Parallelization | Decode speed | Time to first token |
| Two M5 Ultra Mac Studios | Pipeline | 15.74 tokens per second | 8.4 seconds |
| Two M5 Ultra Mac Studios | Tensor | 24.99 tokens per second | 8.2 seconds |
| Four M5 Ultra Mac Studios | Tensor | 36.73 tokens per second | 4.8 seconds |
The two-machine results differ with the parallelization method: Apple reported 15.74 tokens per second with pipeline parallelization and 24.99 with tensor parallelization. Its four-machine tensor-parallel result was 36.73 tokens per second. Each figure belongs to the same stated context and output conditions.
What the distributed setup requires
Apple describes distributed inference over RDMA on Thunderbolt 5. Its setup uses macOS 26.2 and MLX v0.30.1 or later; Apple’s two-machine configuration links each pair with two Thunderbolt cables and uses the jaccl communication backend. Those requirements are part of the specific setup Apple describes, rather than a claim that any two Macs can run the model together. (Apple’s technical report)
How to access the weights
Meta lists Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models as places to download the weights. For hosted API access, it names Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Meta releases the model under the Muse Model License and says many developers can use it for free under limited commercial terms. (Meta’s announcement)
