Meta releases open-source AI model Muse Glimmer

Meta has released Muse Glimmer, a slimmed-down “open source” AI model that can run on a single PC, the company said today.

The model is small enough to fit on one GPU, making it ideal for local tasks like scheduling or file management. Meta calls it a balance between capability and limited hardware. The company is sharing the AI’s “weights” — the decision-making rules it uses — for free on Hugging Face, along with developer guides. Optimized versions are expected soon on sites like llama.cpp, so users can get a working AI agent within minutes.

Muse Glimmer performs strongly for its size, Meta claims, with high success on tests like DeepSearch QA, MCP-Atlas and SWE-Bench, which checks coding and debugging skills. It handles tool use, multi-step reasoning, failure recovery and multimodal input, and supports scaffold compatibility with systems like OpenClaw. The model was trained on data from more than 100 languages.

Meta is positioning the release as part of a push to distribute AI widely rather than centralize it. CEO Mark Zuckerberg wrote that giving people the ability to run models locally spreads power more evenly. The move mirrors efforts from Chinese rivals such as DeepSeek, which also offer locally runnable models with open licenses. Meta’s Muse Spark series has lagged behind rivals from OpenAI and Anthropic, so this release looks like an attempt to win over users who value local control and open access. That approach has gained momentum lately, highlighted by NVIDIA’s new Open Secure AI Alliance, formed after a rogue attack on Hugging Face by an unreleased OpenAI model.