Meta Releases 816B Muse Spark Open Model That Runs on Mac Studios 

Meta on Thursday released the weights for Muse Spark Open, a natively multimodal model with 816B total parameters (42B active) and a 1M-token context window, which it calls the most capable open-weight model from a US company. The model is available today through API platforms at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, with weights downloadable from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models.

The release gives developers a large Meta model they can host and fine-tune themselves, days after Meta kept its stronger Muse Wonder Full model off its API.

Meta built the launch around running the model locally, in partnership with Apple. It says Muse Spark Open is the first model in its class to run on a desktop setup of two M5 Ultra Mac Studios using Apple's MLX framework.

That still takes serious hardware. In Apple's write-up, the model would require more than 1.6TB of RAM in its native precision (BF16) for the model parameters alone, and its 4-bit NVFP4 version takes about 493GB, too much for one 512GB Mac Studio once inference memory is added. At an 8,192-token context, Apple measured decode speeds of 15.7 tokens per second on two machines in pipeline mode, 25.0 on two in tensor-parallel mode and 36.7 on four. Apple did not give a cost for the setup.

The model also runs on leading inference engines like SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth.

The terms are narrower than "open" suggests. Meta is releasing the weights under the Muse Model License with limited commercial terms. According to MarkTechPost, the license permits use, fine-tuning, distillation and redistribution, but products must display "Built with Muse Spark Open." Meta's post did not say which commercial uses fall outside the free terms.

Independent reviewers are split on the "most capable" claim. MarkTechPost rated it strong on agentic tool use; mid-pack among large open models on coding. The New Stack said it is on par with or ahead of some leading Chinese open-weight models depending on the benchmark, and named Thinking Machines' Inkling and Nvidia's Nemotron 3 Ultra as its closest US competitors.

Meta says the open version has additional safety post-training to make it more robust against misuse as a standalone model without system guardrails, and that it met the thresholds our framework requires for release. As with Muse Wonder, the full evaluation report is not yet out; Meta says it will be available soon.

Muse Spark Open is the open-weight version of Muse Spark, the first model from Meta Superintelligence Labs. Meta's previous major open-weight family, Llama 4, was released on April 5, 2025.