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

Meta Platforms Inc. unveiled Muse Glimmer, an open-source language model capable of running on personal computers and Macs. The company also shared a detailed essay from Chief Executive Officer Mark Zuckerberg, addressing AI risks, open-source model regulations, and related topics. Muse Glimmer boasts 30 billion parameters, but Meta's engineers optimized its size to under 20 gigabytes of RAM using techniques like quantization, a method that compressed each weight into four bits without significant loss of performance on certain tasks.

The model operates through a two-step process: a less advanced "drafter" model generates an initial response, which is then verified, refined, and presented to the user – a technique known as speculative decoding. This process is faster than Muse Glimmer generating its own responses. Trained on data from Meta's Muse Spark series, the model has been further refined through two additional training runs.

One focused on handling lengthy prompts and reasoning tasks, while the other made Muse Glimmer better at powering AI agents. Unlike some models that stop if they encounter an obstacle, Meta trained Muse Glimmer to retry tasks it couldn't complete on the first attempt. Users can control the model's "reasoning strength" settings to dictate how much time and computing power it spends on each task.

In benchmark tests, Muse Glimmer outperformed comparable models like Gemma4-31B and Qwen3.6-27B on half of the benchmarks, excelling in areas like online research, code generation, and scientific chart analysis. The launch of Muse Glimmer follows a year since Meta's last open-source AI model. Zuckerberg's essay also discusses the economic benefits of AI, the technology's risks, and calls for reduced regulatory barriers to open-source model development.

He urged the U.S. government to ease such obstacles and suggested adopting governance structures to define AI safety criteria, enabling Meta to evaluate future models and determine their release.