Summary
On October 15, 2026, Meta Superintelligence Labs released Muse Spark Open, the open-weights version of its 816B-parameter sparse mixture-of-experts model (about 42B active per token) trained for long-horizon agentic work such as coding and customer service. It accepts text, image, video, and audio with a 1M-token context, ships BF16 and NVFP4 checkpoints on Hugging Face, and launched with hosted inference on Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, and Together AI.
What changed
Meta published open weights for Muse Spark Open (816B total / ~42B active MoE, multimodal, 1M context) under the Muse Model License, with BF16 (~1.6 TB) and NVFP4 (~530 GB) checkpoints on Hugging Face and day-one hosting across Baseten, Fireworks, Groq (limited preview), Modal, Nebius, OpenRouter, and Together AI, plus listings in Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models.
Why it matters
A frontier-scale, multimodal open-weight model purpose-built for agents, distributed across every major inference provider on launch day, intensifies competition with Chinese open-weight models and gives builders a self-hostable alternative to closed agent models, though the custom Muse Model License sets terms teams must check.
Evidence excerpt
Muse Spark Open is the open-weights version of Muse Spark, an 816B-parameter sparse mixture-of-experts model from Meta Superintelligence Labs, trained for long-horizon agentic work. Hosted inference: Baseten, Fireworks AI, Groq (limited preview), Modal, Nebius, OpenRouter, and Together AI.