Meta releases an open model for local AI agents

Meta releases an open model for local AI agents

The 30-billion-parameter system runs on a single consumer graphics card and supports multi-step tasks without needing the cloud.

GP
Giulio Prisco
Aug 12, 2026
2 min read

Meta has released Muse Glimmer, a language model with 30 billion parameters. The model is designed for agentic work, meaning it can plan and carry out multi-step tasks on its own, such as managing files, writing code, or calling external tools. Its weights are freely available under the Apache 2.0 license so anyone can download and run them.

Unlike most large models that require remote servers, Muse Glimmer is built to operate entirely on a personal computer or laptop that has one ordinary consumer graphics card. This allows it to function with or without an internet connection. The model accepts both text and images through a dedicated perception component, supports more than one hundred languages, and can adjust how much reasoning effort it applies depending on the task.

Training followed three stages. First the model learned by matching the outputs of a larger teacher model called Muse Spark, a process known as distillation in which a smaller model absorbs knowledge from a bigger one. Next it received longer training focused on agent-style work. Finally it was refined with supervised examples and reinforcement methods that reward successful completion of complex jobs.

Performance and practical speeds on local machines

On standard agent benchmarks the model matches or exceeds other systems of similar size. To fit on everyday hardware the full-precision version, which would need more than 55 gigabytes of memory, is compressed through quantization. Quantization lowers the numerical precision of the model’s internal values so the language portion occupies less than 20 gigabytes while still leaving room for temporary working memory and image handling. A small companion network then proposes several tokens at once; the main model checks them in parallel. This technique, called speculative decoding, speeds up generation without changing the final quality.

The result is a responsive local agent that can diagnose failed tool calls and try again, maintain long coherent plans, and operate across common software frameworks. Weights are already available for download, and support for popular local-running tools is expected shortly.

About the Writer

More from Mindplex

Keep reading

Three more ideas worth your time.

Browse News

Discussion

Join the discussion

Sign in to share a response with the community.

Type @ to mention someone Type / or use + to add a block Highlight text, then choose Link
Loading editor

Comments cannot be edited after posting because they become part of the reputation record. Give yours a quick review first.