Meta unveiled Muse Glimmer on Monday, introducing a 30-billion-parameter model designed for agent work that can run on a laptop. The company made the model’s weights available – the trained numbers that form the model itself, allowing anyone to download and execute it on their own machine – on the model repository Hugging Face, under an Apache 2.0 license that is notably permissive, without the usage restrictions Meta imposed on its Llama releases.
The model is targeted at a specific and increasingly competitive segment: AI that operates on personal hardware rather than relying on external cloud services. Google has Gemma, Alibaba offers Qwen, while Mistral and DeepSeek provide compact open models. Meta is entering a market it arguably established but then neglected.
Overview
Glimmer was developed from Muse Spark, Meta’s cutting-edge model, using a technique known as distillation: training the smaller model on the output of the larger model until it can replicate what the bigger system already knows.
However, there is the challenge of fitting the model. A 30-billion-parameter model at full precision demands over 55GB of memory, which no consumer graphics card possesses. Meta compressed the model’s components to about a quarter of their standard precision, reducing it to under 20GB – small enough to accommodate other necessary processes within a 24GB or 32GB card. The company claims that the compression has minimal or no impact on critical tasks.
Speed is enhanced through a technique called speculative decoding. Traditional models write one word at a time, with each word waiting for the previous one, resulting in slow responses for lengthy answers. Meta pairs Glimmer with a swift companion model that predicts complete text segments, allowing the primary model to validate the predictions collectively and retain the correct ones. This method is effective because validating an answer is quicker than generating one. Meta reports that the model is 3.1 times faster on an RTX 5090, 1.8 times on an M5 Max, and 1.5 times on an M4 Max.
Meta has positioned Glimmer against Google’s Gemma4-31B and Alibaba’s Qwen3.6-27B, showcasing impressive outcomes on tests assessing a model’s ability to complete multi-step tasks end-to-end – such as resolving genuine bugs in codebases, integrating external tools, and recovering from failures. These results are based on Meta’s internal assessments, emphasizing the importance of independent testing. As of Monday, the model is available for public access.
Meta CEO Mark Zuckerberg has announced that a version of Spark will be released in the upcoming weeks, followed by larger models. Downloadable models are more cost-effective and customizable compared to rented models, with many of the top downloadable models originating from Chinese developers like DeepSeek, Alibaba, and Moonshot. Zuckerberg argues that American labs face limitations due to data restrictions that foreign competitors do not encounter, emphasizing that barring foreign models is not the solution.
Zuckerberg advocates for safeguarding distillation as a policy, highlighting that learning from observable data is crucial. Glimmer, being a distilled model released on the same day, serves as an example of this principle in action.
Additionally
The unveiling of the model was accompanied by a 6,500-word essay titled “The Future is for Everyone,” asserting that advanced AI should be accessible to individuals rather than concentrated in a few entities, challenging the notion that extreme centralization of AI power is the only safe approach.
Meta also disclosed a $1 billion fund for communities hosting its data centers, addressing the resistance faced by local communities in the establishment of AI infrastructure. The essay cites the positive impact in Richland Parish, Louisiana, where teachers received bonuses from the tax revenue generated by Meta’s construction projects.