Meta releases open-weights Muse Glimmer model with 30B parameters

Meta Platforms has put out Muse Glimmer, an open-weights language model built to run on personal computers. The company paired the release with a long essay from Mark Zuckerberg that tries to do two things at once: defend open model development and warn about the risks that come with it. That combination is the real story here. This isn’t just a model dump; it’s Meta signaling that it wants back into the open-source conversation in a big way.

Muse Glimmer has 30 billion parameters, which would normally mean roughly 55 gigabytes of RAM. Meta says it cut the model down to under 20 gigabytes through optimization work, including quantization that compresses weights into four bits. The result is practical enough to run on personal computers and Macs with a single consumer-grade graphics card. Meta also says the quantization caused “minimal to no degradation on agentic tasks.”

The model itself uses a two-step setup for answers. A smaller “drafter” model makes an initial attempt, and Muse Glimmer then checks, refines and returns the final response. Meta says that speculative decoding is faster than having Muse Spark generate answers alone. Under the hood, the company trained Muse Glimmer on data made by its proprietary Muse Spark series, then ran two more training rounds: one to improve long prompts and reasoning, and another to make the model better at AI agents. Meta also says it trained the model to retry tasks it fails the first time, and users can adjust its “reasoning strength” depending on how much time and compute they want it to spend.

Meta tested Muse Glimmer on two dozen popular AI benchmarks. The company says it beat the similarly sized Gemma4-31B and Qwen3.6-27B on half of them, including tests for online research, code generation and scientific chart analysis. That’s a neat line on the release post, but the more important part may be the timing: this is Meta’s first open-source AI model in more than a year, and Zuckerberg says the next one is coming “soon.”

His essay also pushes hard on policy. He wants the U.S. government to lower barriers for open-source model development, says Meta’s board is building a governance structure to define AI safety criteria, and argues that frontier labs should give governments early access to models while they’re still being trained. The pitch is familiar Meta: open the doors, but keep a lock on who gets to decide whether the house is safe.