Meta released the open weights for Muse Spark Open on October 15, 2026, describing it as the most capable US open-weight model to date. The model is a sparse mixture of experts with about 816 billion total parameters and roughly 42 billion active per token. It takes text, images, audio and video as input, produces text, and supports a one-million-token context window. Meta Superintelligence Labs positions the release for coding and long-horizon agent work rather than as a general chat demo.
Alongside the Hugging Face checkpoints, Meta claims a first-of-class desktop path: a pair of Apple M5 Ultra Mac Studios running the newest MLX stack. Unified memory on each Studio can expose up to 512 GB to the GPU, so two boxes can hold the weights. Zuckerberg and Apple CEO John Ternus both posted about that Mac Studio route on launch day, and The Verge said the firms had collaborated for several months.
Weights are posted under the Muse Model License with limited commercial terms that Meta says let many developers use the model free of charge. Meta published BF16 and NVFP4 checkpoints on Hugging Face under the meta-models organisation, and listed hosting on Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Day-one inference partners named in the research post include Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
Serving paths named by Meta include SGLang and vLLM among others, plus Apple MLX and llama.cpp for local stacks, with silicon notes for NVIDIA, AMD, Intel and Apple. Helion kernels inside SGLang are cited for GPU speedups on NVIDIA and AMD. Further adaptation can use PyTorch TorchTitan.
Meta framed the open release as part of bringing advanced models to people who want to run them locally, while adding extra safety post-training because open weights remove some platform guardrails. It said the model met release thresholds under its Meta Superintelligence Scaling Framework and pointed readers to the model card for detailed evaluations, with a fuller evaluation report said to follow.
Independent coverage and partner posts confirm that weights are live and that inference providers already advertise Muse Spark Open. Hugging Face published a practical walkthrough the same day covering architecture numbers, chat template notes and fine-tuning paths. Groq and Together AI both announced limited preview or live hosting. Those secondary posts do not independently re-measure Meta's capability claims; they mainly establish availability.
The open weights sit beside Meta's earlier Muse line, including Muse Wonder models covered earlier in October and the smaller Muse Glimmer dense model. Muse Spark Open is a distinct weight drop: a large multimodal MoE aimed at agentic coding, with an unusual desktop path through Apple silicon. Pricing for hosted APIs, exact commercial limits in the Muse Model License and independent third-party leaderboard results beyond Meta's own write-up were not fully settled in the materials reviewed for this package.