Meta has released Muse Glimmer, a new 30B parameter language model under the permissive Apache 2.0 license. The company claims the model is optimized for local use, a shift from its earlier Llama licenses, which developer Simon Willison described as "janky." The release marks a notable step toward more flexible open-weight models that users can run on their own hardware.
A Model Built for the Desktop
Muse Glimmer's 30B parameter count sits in a sweet spot for local execution. Simon Willison, a blogger and developer known for creating Datasette and various LLM plugins, tested the model on his own machine. He generated a pelican image using LM Studio's 18.16 GB version of the model. That file size fits comfortably within the memory of many modern systems.
Willison's machine has 128GB of RAM, but he notes the model is practical for more modest setups. He says the model leaves plenty of space for other applications on machines with 32GB+ RAM. That is a key advantage for developers who want to run a model alongside browsers, editors, and other tools without constant memory pressure.
Testing on Real Code
Willison did not stop at image generation. He put Muse Glimmer through a coding test using his llm-coding-agent plugin on a fresh checkout of Datasette. The test involved a prompt and a long transcript of tool calls to explore the codebase. This kind of hands-on evaluation gives a clearer picture of how the model handles multi-step reasoning and code navigation.
The results, according to Willison, imply the model performs well in coding tasks. He did not provide benchmark scores, but his practical test suggests the model can handle real-world development workflows. For a 30B model running locally, that is a meaningful signal.
Why the License Matters
Apache 2.0 is a permissive open-source license, and its use here is significant. Meta previously relied on Llama licenses that imposed restrictions on commercial use and large-scale deployments. Those terms frustrated many developers and companies. With Apache 2.0, Muse Glimmer can be used, modified, and redistributed with fewer constraints.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Willison expressed approval of this licensing choice. The shift could encourage broader adoption of Meta's models in open-source projects and commercial products alike. Open weights models allow local execution, which appeals to users concerned about privacy, cost, or network dependence.
Practical Implications for Developers
The 30B size is a deliberate trade-off. Smaller models run faster but often lack reasoning depth. Larger models are more capable but demand expensive hardware. Muse Glimmer aims to balance those factors, and Willison likes this size of model because it leaves plenty of RAM for other applications.
For developers, this means running a capable assistant alongside their usual tooling without dedicating an entire machine to it. The 18.16 GB file size also makes downloading and managing the model straightforward. On a machine with 32GB or more of RAM, users can keep Muse Glimmer loaded while working on other tasks.
A Link Post with a Sponsor Option
Willison's write-up, posted on 10th August 2026, is a link post rather than a full tutorial. He offers a sponsor option for $10/month for a curated email digest. That digest likely covers developments like this one, giving subscribers a regular roundup of tools, models, and techniques.
The post itself is concise, but the testing details add substance. Willison's approach of combining image generation and coding tests gives readers a practical sense of what Muse Glimmer can do. His approval of the Apache 2.0 license and the model's local-friendly size suggests he sees it as a useful addition to the open-weights ecosystem.