Meta released Muse Spark Open, a US open-weight model. It has 816B total parameters and a 1M-token context window. It excels at coding, agentic work and long-horizon tasks. It runs on two Apple M5 Ultra Mac Studios with MLX. Weights are available on Hugging Face, SageMaker, Google Cloud and others. It works with SGLang, vLLM, TensorRT-LLM and more. API access is live via Baseten, Fireworks, Groq and others. It passed Meta’s safety evaluation for release. The model uses a limited commercial license. Developers can fine-tune it with PyTorch.
What changed
No US open-weight model ran on desktop hardware with this capability.
What it unlocks
Running a leading open-weight model on two Apple M5 Ultra Mac Studios for coding and agentic tasks.
- 816B total parameters
- 42B active parameters
- 1M-token context window
- 512 GB memory per Mac Studio
What you need to act on it
- two Apple M5 Ultra Mac Studios
- MLX framework
Sources