Meta and Apple run 816B Muse Spark Open on two M5 Ultra Mac Studios as Meta returns to open weights

Meta is back in open weights. Muse Spark Open is an 816-billion-parameter mixture-of-experts model that activates 42 billion parameters per token, aimed at software engineering and long-horizon agent tasks. It is the open member of the Muse family, which has replaced the Llama brand. The weights are hosted on Hugging Face under the Muse Model License, which permits free use with limited commercial terms. Meta recommends TorchTitan for customization and Apple's MLX for running on Apple Silicon.

The Apple partnership is the technical headline. The two companies showed the full model running natively across two M5 Ultra Mac Studios at more than 36 tokens per second. That is possible because the MoE design only activates 42B parameters per token, and the Mac Studios' large unified memory can hold the full set of weights across two machines. It is a real milestone for local frontier-class inference, though not a cheap one: estimates of the two-Mac setup range from about $10,800 to $20,000.

The release came days after Meta debuted its closed frontier models, Muse Wonder and Wonder Full, priced at $4/$20 per million tokens through the Meta Model API. That split drew skepticism: Meta is open-sourcing the weaker model and keeping the best one closed. Testers report strong image-to-HTML transcription and tool calling roughly matching DeepSeek V4.1, though outputs lag Meta's closed tiers. SemiAnalysis's "American open weights are back" line, amplified by Meta, captures the competitive frame against Chinese open models.

Open questions: the license's commercial limits, whether the community fine-tunes it heavily as it did Llama, and whether Meta keeps shipping open versions as Muse Wonder evolves. Separately this week, Meta removed business-terms language that had raised scraping and AI-training concerns.