Meta Drops 30B-Param Muse Glimmer Weights, Says Open Source or Bust 🪨

Meta has released the weights for its Muse Glimmer open-weight AI model, a 30-billion-parameter dense model that can be run locally, CEO Mark Zuckerberg announced in an X post. The company also disclosed plans to soon release the weights for Muse Spark 1.2, which Zuckerberg described as its "latest and most powerful frontier model." "Meta is a strong supporter of open source and I'm proud of these releases," Zuckerberg said.

The release comes amid intensifying competition from Chinese open-weight AI models. Zuckerberg reiterated the case for open-source development, framing it as central to the United States' effort to maintain leadership in the AI race. Meta was among the technology firms that signed an open letter warning the Trump administration against restrictions on open-weight model development.

Open-weight releases differ from full open-source releases in that the model weights are published for external use and fine-tuning, while the underlying training data and full codebase may remain restricted. Muse Glimmer's local-runnable design places it in the same operational class as other recently released dense and mixture-of-experts models that developers can deploy without relying on a centralized API.

Meta has not disclosed a specific release date for Muse Spark 1.2 weights. The company previously launched versions of its Llama model family as open-weight releases, a strategy that has influenced how startups, enterprises, and researchers select foundation models for downstream applications.

Zuckerberg's post did not specify licensing terms, parameter counts beyond the 30-billion figure, or benchmark results for Muse Glimmer. No direct comment from Chinese AI model developers was included in the announcement.