Meta releases Muse Spark Open weights
The short version
Meta has released Muse Spark Open weights for a multimodal MoE aimed at coding and agentic work, with 816B total parameters, 42B active parameters, and a 1M-token context window.
What changed
Meta released Muse Spark Open weights through its official Hugging Face collection and listed several hosted API paths.
What it means for you
Teams that need control of model deployment boundaries can evaluate the weights or a hosted path, but should verify access, hardware, and cost first.
Meta has released weights for Muse Spark Open. The company describes it as a multimodal mixture-of-experts model for coding and agentic work, with 816 billion total parameters, 42 billion active parameters, and a one-million-token context window.
The official entry point is Meta's Hugging Face collection. Meta also lists hosted availability at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Its build guide names TorchTitan, SGLang with Miles, NVIDIA NeMo AutoModel, and Hugging Face TRL as post-training paths.
For teams that need control over runtime, infrastructure, and data boundaries, the release creates both an open-weight and a hosted evaluation path. Meta's desktop demonstration uses two Apple M5 Ultra Mac Studios, so it should not be treated as evidence that ordinary hardware can run the model cheaply. Test latency, cost, context use, and tool permissions in a constrained read-only workflow before production integration.
Related tools
Fact check
Coverage timeline
- Primary sourceMetaIntroducing Muse Spark Open