Meta releases Muse Spark Open, an open-weight ~816B-parameter multimodal mixture-of-experts model, with Hugging Face support

Meta has published open weights for Muse Spark Open, a mixture-of-experts model with about 816B total parameters and about 42B active per token (8 of 256 experts routed). It accepts text, image, audio and video input, outputs text only, supports up to 1M tokens of context and is tuned for agentic coding. Checkpoints are on Hugging Face in BF16 (about 1.5 TB) and NVFP4 for Blackwell GPUs, with Transformers support and llama.cpp, vLLM and SGLang recipes. It follows the small dense Muse Glimmer released in August.

Developers, researchers and enterprises that want to self-host, fine-tune or audit a very large multimodal model now have a US-origin option with mainstream tooling support from day one. It gives enterprises with data-residency or cost concerns an alternative to closed APIs, and it puts pressure on other open-weight and closed-model vendors. In practice the audience is narrow: running it needs several H200 nodes or a B200 node, so most teams will use hosted inference rather than run it themselves. The licence terms are not in the excerpt and will decide how much commercial use is really allowed.

"Most capable US open-weights model" and "frontier" are Meta's own framing, and the excerpt gives no benchmarks or independent evaluations to support them. The 1M-token context figure is a maximum, not evidence that the model uses long inputs well. "Open" here means open weights, not necessarily open training data, code or an unrestricted licence.

1. 01Independent benchmark results within the next 2 to 4 weeks (agentic coding suites such as SWE-bench-style tests, plus long-context and multimodal evaluations) compared with the strongest existing open-weight models and with closed models.
2. 02The licence text and acceptable-use terms, plus any restrictions on commercial use, large-scale deployment or fine-tuned derivatives.
3. 03Adoption signals: Hugging Face downloads and derivative fine-tunes, availability on third-party inference providers and cloud platforms, and reported per-token prices and throughput on vLLM and SGLang.