Meta
Muse Spark Open
downloadable weights for multimodal agents

Key facts
- 15 Oct 2026
- Released
- 816B total / 42B active
- Parameters
- 1,048,576 tokens
- Context
- Text
- Output
Muse Spark Open is Meta's downloadable AI model for answering questions and taking actions through tools. Meta released the weights on 15 October 2026.
Muse Spark Open is Meta’s downloadable AI model for answering questions and taking actions through tools. Meta released the weights on 15 October 2026.
It reads several kinds of input and writes text
The model accepts text, images, audio and video and generates text, according to the Hugging Face technical guide. Its 1,048,576-token window is shared by the inputs and the conversation.
Muse Spark Open has about 816 billion total parameters, with about 42 billion active per token. A mixture-of-experts design selects part of the network for each step. The full model still needs substantial memory, including the inactive experts.
Developers can choose BF16 or NVFP4 checkpoints
Meta’s official collection provides BF16 and NVFP4 versions. Quantisation stores model values at lower precision to reduce memory use. Hardware, runtime and context length determine the full deployment requirements.
Hosted providers also offer access. Their prices and supported features belong to those providers and can differ from a self-hosted deployment.
The Muse Model License sets commercial conditions
Meta publishes the weights under the Muse Model License with limited commercial terms. Its announcement says many developers can use the model free of charge under those conditions. Teams planning a product should read the current licence and access terms before choosing a deployment.
The release gives developers the files needed to adapt and host the model. Its practical value depends on task success, hosting resources and the permissions attached to its tools.
More in Large Language Models
All LLMs →- MetaMuse SparkMeta's proprietary frontier family
- MetaMuse WonderMeta's models for coding and longer tasks
- MiniMaxMiniMax M3.1 Flasha low-cost model for coding agents
- Google DeepMindGemini 4 Argonlonger reasoning, coding and cyber defence
- OpenAIGPT-5.6the flagship from 9 July to 3 September 2026
- OpenAIGPT-6 Astrathe new generation, rolling out from 3 September 2026