Meta released Muse Spark Open on October 15, 2026. It is an open-weight multimodal model from Meta Superintelligence Labs. The model uses a sparse mixture-of-experts architecture with 816B total parameters and about 42B active per token. It supports a context window of up to 1,048,576 tokens. It accepts text, image, video and audio input and generates text. Meta positions it for coding, agent workflows and long-horizon tasks. The weights are available under Meta’s Muse Model License. It can be accessed via a hosted playground, an API or self-hosted inference. Self-hosting requires substantial GPU memory and specialized setups. The model does not generate images, video or audio. Providers may impose their own limits on inputs, context and features.
What it unlocks
Running a 816B-parameter sparse mixture-of-experts model with 1M-token context for coding and agent workflows.
- 816B total parameters
- 42B active parameters per token
- 1,048,576 token context window
- 1.65 TB BF16 checkpoint
What you need to act on it
- compatible runtime and hardware
- acceptance of Muse Model License
Sources