Meta released Muse Spark Open on October 15, 2026. Its announcement makes downloadable model weights available to developers.
What the model actually contains
The model card specifies 816 billion total parameters, 42 billion active parameters and a one-million-token context window. Inputs can include text, images, audio and video; output is text. Meta positions the model for programming and multi-step agent tasks.
Hugging Face’s technical introduction explains the mixture-of-experts design: eight of 256 experts are selected for each token. The important distinction is that inactive experts still occupy memory. A smaller active-parameter count does not turn the full model into a small download or remove its storage requirements.
Local deployment means a serious workstation
Meta’s desktop demonstration uses two M5 Ultra Mac Studios, each with up to 512 GB of unified memory. This is a two-computer installation, not a claim that an ordinary laptop can run the complete model.
The relevant hardware concept is explained in the MLX unified-memory documentation: CPU and GPU share a memory pool, avoiding the need to move an array between separate memory locations for every change of processor. That explains the architecture, not the speed of any particular application.
The MLX LM project supplies language-model inference and fine-tuning tools, Hugging Face integration and distributed execution. Its documentation also describes prompt caching and the memory-versus-speed trade-off when processing long prompts in smaller chunks.
Open weights come with conditions
The Muse Model License requires attribution and compliance with a usage policy. Model-as-a-service businesses whose aggregate revenue with affiliates exceeds US$20 million in a consecutive 12-month period need a separate Meta agreement. The specified exception covers internal use and products that are not model-as-a-service offerings. Calling this unrestricted open source would be inaccurate.
How to read the performance claims
Meta’s evaluation methodology combines official benchmark results, internal testing and some competitors’ self-reported scores. It also acknowledges that its tools and prompts were not specifically optimized for every rival. These are important qualifications when interpreting the company’s comparative claims.
Analysis: a deployment choice, not a turnkey agent
The distinction between a model and a finished assistant matters here. Teams still need the application around it: tool connections, permissions, error handling and a way to review changes. The release gives them more control over that stack, while leaving its integration in their hands.
Sources: Meta 1 · Hugging Face 1 · GitHub 1 · GitHub 2 · GitHub 3 · Hugging Face 2 · Meta 2.
Comments