Meta's Muse Glimmer is a free 30B AI agent model that runs on a single consumer GPU
Meta has released Muse Glimmer, an open model built to run AI agents locally, with no cloud needed. It handles text and images, plans multi-step tasks and recovers from failures, and it's free under the Apache 2.0 license.
By ScoviAI Team
Key points
- A 30-billion-parameter open model for autonomous agent tasks on consumer hardware
- Compressed to about 4-bit precision so it runs on a single GPU
- Accepts text and images; trained on data in over 100 languages
- Released under Apache 2.0; Meta says it beats Gemma4-31B and Qwen3.6-27B on agent benchmarks
Most AI agents live in the cloud: you send a task to a company's servers and wait for the result. Muse Glimmer, Meta's newest open model, is built to do that work on your own machine.
Meta describes it as "purpose-built for autonomous agentic tasks on consumer hardware."
What Muse Glimmer is
| Details | |
|---|---|
| Size | 30 billion parameters |
| Inputs | Text and images (via a built-in perception encoder) |
| Output | Text |
| Languages | Trained on data from over 100 languages |
| Precision | Quantized to about 4-bit for single-GPU use |
| License | Apache 2.0 |
Quantizing a model means storing its numbers with less precision so it takes up less memory. At about 4 bits, a 30-billion-parameter model can run on a single consumer graphics card instead of a server rack.
What it can do
Meta says Muse Glimmer combines three skills agents need in one model:
- Multi-step reasoning: breaking a task into steps and working through them.
- Reliable tool use: calling apps, files or APIs correctly.
- Failure recovery: noticing when a step fails and trying something else.
Meta claims it delivers state-of-the-art results for its size on agent benchmarks, beating Google's Gemma4-31B and Alibaba's Qwen3.6-27B on tool use, multi-step planning and task completion. These are Meta's own benchmark results and haven't been independently checked yet.
How it fits Meta's model family
Muse Glimmer sits alongside Meta's Muse Spark and Muse Wonder models. The bigger Muse Wonder models are Meta's frontier systems for agents, which we covered in Meta launches Muse Wonder. Glimmer is the small, local, open one. Meta's AI for Good team also highlights it for humanitarian and social-impact work, where cloud access can be expensive or unreliable.
Why it matters
Running agents locally means your files and prompts stay on your device, there's no per-use bill, and the agent keeps working without an internet connection. The catch is hardware: you still need a capable GPU. Laptops such as NVIDIA's RTX Spark machines are being built for exactly this kind of model. And because Glimmer is released under the permissive Apache 2.0 license, companies can build on it freely. New to the idea? Read open-weight AI models, explained.
Sources
ScoviAI Team is the ScoviAI newsroom. We write every article ourselves from the original announcements and link our sources. Spotted a mistake? Tell us and we'll fix it.
ADVERTISEMENT
