Meta releases Muse Spark Open for coding and agent workflows
Downloadable weights bring a million-token multimodal model to developers, under a license with commercial conditions.
THE STORY IN THREE POINTS
- Meta released Muse Spark Open weights on October 15.
- The model has 816 billion total parameters, 42 billion active, and a million-token context.
- The Muse Model License includes commercial conditions.
What Meta released
Meta released Muse Spark Open on October 15, making model weights available alongside access through hosted inference providers. It accepts text, images, audio and video, and generates text. The company positions it for software development and tasks that require several coordinated actions.
The model uses 816 billion parameters in total, activating 42 billion at a time, with a context window of one million tokens. Meta describes support for multiple inference engines and Apple MLX. Its Muse Model License carries commercial conditions; downloadable weights do not mean unrestricted use.
Editorial context: control includes operations
For a development team, hosting a model changes who manages upgrades, capacity and access. Those responsibilities belong in a comparison with an API service, alongside the ability to customize a deployment.
A useful trial could give both options the same repository task and record accepted changes, human review time and total operating cost. A large context window alone cannot establish whether an agent understood the relevant code.
Editorial outlook: measure completed work
The practical question is how consistently the model delivers changes that maintainers can merge. Repeated trials with unfamiliar repositories would reveal more than a polished demonstration. Our newsroom has not independently tested its coding performance or hardware requirements.
Sources: Meta: Introducing Muse Spark Open ↗
