Meta's Muse Spark Open: 816B open weights that run on two Macs

Meta has released the weights of Muse Spark Open, which it calls the most capable US open-weight model to date. It is an 816B-parameter mixture-of-experts model, but only 42B parameters are active per token. It has a 1M-token context window, is natively multimodal, and is tuned for coding and agentic work: long-horizon tasks, complex software engineering and multi-step reasoning.

The surprising part is the hardware. Meta says the full model runs locally on two Apple M5 Ultra Mac Studios with up to 512GB of unified memory each, using MLX. You can download the weights from Hugging Face, SageMaker, Google Cloud Model Garden and Microsoft Foundry. Hosted APIs include Groq, Fireworks, Together, Baseten, OpenRouter, Modal and Nebius. On day one it works with vLLM, SGLang, TensorRT-LLM, llama.cpp and Unsloth. Meta's demos pair it with Muse Code, its coding terminal, to fix bugs and build a full Mastodon client.

Read the fine print. The weights come under the Muse Model License, which has "limited commercial terms", so this is not Apache-style open. The "most capable" claim is based on Meta's own evaluations, and no one has checked it independently yet.

Why it matters: Meta is back to releasing open weights at frontier scale, and a model this size running on desktop hardware brings serious agentic coding a lot closer to running locally.