Meta releases weights for its 816B-parameter Muse Spark Open model
Muse Spark Open combines 816 billion total and 42 billion active parameters with a one-million-token context window, accepting text, images, audio, and video as input.
AT A GLANCE
Key points
- 816B total parameters, with 42B active per token.
- One-million-token context with text, image, audio, and video inputs.
- BF16 and NVFP4 weights are available under the Muse Model License.
- Local NVFP4 deployment requires at least 700GB of VRAM.
What is changing?
Meta released the weights for Muse Spark Open through Hugging Face on October 15. The mixture-of-experts model has 816 billion total parameters and activates 42 billion per token. It offers a one-million-token context window, accepts text, images, audio, and video as input, and produces text output.
The model card lists long-horizon reasoning, coding agents, tool use, and multistep task completion among its primary use cases. BF16 and NVFP4 weights are available, with support for Transformers, vLLM, SGLang, TensorRT-LLM, Unsloth, llama.cpp, and MLX. Meta says local NVFP4 deployment requires at least 700GB of VRAM.
The weights are distributed under the custom Muse Model License, so “open weights” is more accurate than “open source.” Meta also stresses that the model is not a standalone safety layer and recommends additional safeguards tailored to the deployment context when tools are connected.
Explore the original sources
Source date: 15 October 2026.
Published by AI Radar: 16 Oct 2026, 18:05 (Istanbul, UTC+3).
FOLLOW THE TOPIC
Topics and tags
Share your discovery.
Send this story to a friend.
Reader comments
What do you think about this development? Share your experience or question. Comments are published after editorial approval.
Your name/nickname and comment become public if approved. Read: Privacy notice.
Loading comments…