Meta Muse Spark Open: 816B MoE open weights, 42B active, 1M context

Meta released Muse Spark Open, the open-weight version of Muse Spark, on October 15, 2026. Meta's developer blog describes an 816B-parameter sparse mixture-of-experts model that activates about 42B parameters per token. The Hugging Face model card lists 256 experts with 8 active per token; the Hub's safetensors metadata counts 822B parameters. Context is 1M tokens. Input is text, image, video and audio, and the card lists the output as text. Meta publishes BF16 (about 1.6 TB) and NVFP4 (about 530 GB) checkpoints. Partner builds are AMD FP8 and MXFP4 (Quark), Intel MXFP4 and INT4 (GPTQ), and Unsloth GGUF down to about 2-bit. The blog lists SGLang, vLLM, TensorRT-LLM, Transformers, llama.cpp, MLX and Unsloth as supported at launch, and seven hosted providers, including Fireworks AI, Together AI and OpenRouter. Memory guidance differs by source. The model card says NVFP4 needs at least 700GB of VRAM and BF16 needs 2TB. The blog says MLX can serve the NVFP4 checkpoint on two Mac Studios with M5 Ultra and 512GB each, or four Mac Studios with 256GB each. For the smallest Unsloth GGUF, the blog's text says about 300 GB of combined RAM and VRAM, but its table says a 256 GB setup. The licence is Meta's Muse Model License. The LICENSE.md on Hugging Face requires a "Built with Muse Spark Open" notice. If the licensee or its affiliates offer AI models to third parties as a service, and their combined gross revenue exceeds $20M in any 12-month period, the licensee must notify Meta and enter a separate written agreement before offering the model in that service. Internal use, research and the licensee's own non-service products are excluded. Downloads are gated with manual approval. The form asks for name, date of birth, country, affiliation and job title.