Muse Glimmer is an open-weight agentic AI model from Meta Superintelligence Labs: 30 billion parameters built to run autonomous agents locally on a Mac or a PC with one consumer GPU. The weights download free from Hugging Face under an Apache 2.0 license, commercial projects included. It marked Meta's first fully open release since the Muse family went proprietary. Over 100 languages are supported.
Muse Glimmer is a dense multimodal model distilled from Muse Spark, Meta's proprietary frontier model: the small one learns from the big one through knowledge distillation, then goes through agent-heavy training and a reinforcement learning phase. The result plans across multiple steps, chains tool calls, writes code and reads images, with a context window above 131,000 tokens.
In practice, you hand it a repository cleanup or a folder to sort, switch off the Wi-Fi, and it keeps working. A tool call fails midway? The model retries instead of giving up, a behavior Meta trained for on purpose. That makes it a solid base for autonomous AI agents that never send a single byte to the cloud.
On agent benchmarks, Muse Glimmer beats Gemma4-31B and Qwen3.6-27B, its two closest rivals by size, scoring 75.5 on MCP Atlas. It also leads on DeepSearch QA, Gaia2 and SWE-Bench Pro. The flip side: Qwen3.6-27B keeps the edge on computer use and terminal work.
| Benchmark | Muse Glimmer score | Leader |
|---|---|---|
| MCP Atlas (tool orchestration) | 75.5 | Muse Glimmer |
| DeepSearch QA (research) | 74.6 | Muse Glimmer |
| SWE-Bench Pro (coding) | 51.2 | Muse Glimmer |
| AIME 2026 (math) | 94.7 | Muse Glimmer |
| OSWorld-Verified (computer use) | 65.9 | Qwen3.6-27B (75.6) |
Muse Glimmer fits into a 24 to 32 GB memory envelope once quantized: 4-bit compression brings its footprint from roughly 55 GB down under 20 GB, and a small speculative-decoding module called DFlash speeds up generation considerably. Ollama, LM Studio, llama.cpp and vLLM supported it from day one.
Downloads go through Hugging Face (the main GGUF file weighs close to 17 GB, so brew a coffee first). No 24 GB card at hand? Together AI, Fireworks AI and OpenRouter serve the model online with pay-as-you-go billing. On AMD hardware, early tests reached up to 53 tokens per second on a Radeon AI PRO R9700.
Yes, Muse Glimmer's weights download free from Hugging Face, with no per-token fees or subscription. The real cost shifts to hardware: a machine with at least 24 GB of memory, plus electricity. Hosted versions on Together AI, Fireworks AI or OpenRouter are billed per usage instead.
Apache 2.0 licensing permits commercial use, modification and redistribution of the model, including after fine-tuning on your own data. It is among the most permissive licenses in the field, the same one OpenAI picked for its gpt-oss models.
OpenAI's gpt-oss models are text-only mixture-of-experts designs, while Muse Glimmer is a dense model with native vision, trained end to end around the agent loop and shipped with quantized variants tuned for 24 GB machines. Both carry the Apache 2.0 license and adjustable reasoning.
No. The model takes text and images as input, but no audio. Video is not genuinely optimized either: it gets processed frame by frame. An on-device agent meant to listen to meetings or watch a stream will require a separate transcription layer upstream.
Verdict: Developers who already own a 24 GB card get a full agent to host at home, sensitive data included. Heavy workloads may still call for a large cloud model as a sidekick, but for coding, sorting files and private automation, Muse Glimmer earns its place on your own machine.
