Meta released Muse Spark Open on October 15, giving developers downloadable weights for a multimodal model built for coding and multi-step agent work. It has 816 billion total parameters, with 42 billion active per token, and a one-million-token context window. It accepts text, images, audio, and video, and generates text.
Open weights bring deployment choice
Teams can customize and host the model themselves, or access it through providers including Fireworks, OpenRouter, and Together AI. Unlike the preview access described in our coverage of Meta’s Muse Wonder models, this release puts downloadable weights at the center of the deployment story.
Meta also describes local deployment using two M5 Ultra Mac Studios with MLX—not an ordinary laptop setup. The active parameter count describes per-token computation; serving the model still requires loading the full weights, so it does not imply the memory footprint of a 42-billion-parameter model.
Why it matters for builders
Open weights give builders another option for agents that inspect screenshots, work across codebases, and call tools. Self-hosting also makes infrastructure, reliability, and operating costs the developer’s responsibility.
For a studio like SunMarc App Labs, a useful evaluation would be a bounded workflow: inspect an app screenshot, propose a fix, and run checks in a sandbox before human review. This is an evaluation idea, not an announced SunMarc feature or deployment.
The question is whether the model completes that job reliably at a sensible total cost—not simply whether it can be downloaded. Measure successful task completion, retries, test results, and infrastructure use alongside model quality.
Open weights are not unrestricted use
The Muse Model License includes attribution and commercial conditions. Teams should review the actual license and usage policy for their intended deployment; downloadable weights do not remove those obligations or make the release unrestricted open-source software.