Meta · 2026-08-10 · major
Muse Glimmer — Meta's 30B open agentic model runs on one consumer GPU
Muse Glimmer is Meta's 30B open-weight agentic model, released under Apache-2.0 to run locally on a single consumer GPU. It scores 75.5 on MCP Atlas and 51.2 on SWE-Bench Pro, ahead of Gemma4-31B and Qwen3.6-27B on both.
A 30B open-weight agent model from Meta that plans, calls tools and recovers from its own errors on a single desktop GPU.
Quick facts
| Maker | Meta |
|---|---|
| Parameters | ~29.6B dense, incl. 1.8B vision encoder |
| License | Apache-2.0 |
| Context window | 131,072 tokens |
| Runs on | 24GB VRAM at 4-bit (under 20GB) |
| Availability | Open weights on Hugging Face; Ollama, LM Studio, vLLM, Together AI, Fireworks AI |
Benchmarks
What is it?
Muse Glimmer adds an always-on local option to Meta's Muse family: a 30B open-weight model built to act as an agent rather than just chat. It plans a task, calls tools, checks its own results and retries when a step fails. The model reads images as well as text, and Meta ships it under Apache-2.0 with the weights on Hugging Face.
How does it work?
Training started as logit distillation from the larger Muse Spark, followed by mid-training on long-context agent traces and reinforcement-learning post-training. The 52-layer dense transformer repeats a [Local, Local, Local, Global] attention pattern over 2,048-token sliding windows and uses grouped query attention with 32 query heads and 2 KV heads, plus a ~1.8B ViT-G/14 encoder for vision. Failure recovery is native: Muse Glimmer diagnoses schema or runtime errors during a tool call and retries the step itself.
Why does it matter?
Local agents have been blocked by size, not ambition — anything good enough to run a long tool-using loop needed a datacenter. At 4-bit the Muse Glimmer weights drop under 20GB and fit in 24GB of VRAM, and speculative decoding makes generation up to 3.1x faster, so a laptop or desktop can run an agent overnight with no cloud round-trip and no per-token bill. Apache-2.0 also means teams can ship it inside their own products.
Who is it for?
developers building local or private AI agents
Frequently asked questions
- What hardware do you need to run Muse Glimmer?
- Muse Glimmer targets consumer machines. The 4-bit quantized weights come in under 20GB and fit inside 24GB of VRAM with minimal loss in quality, so a single desktop GPU or a Mac with enough unified memory is enough. Speculative decoding speeds generation up to 3.1x, which is what makes a long agent loop practical locally.
- Can you use Muse Glimmer in a commercial product?
- Yes. Meta released Muse Glimmer under Apache-2.0, a permissive license that allows commercial use, modification and redistribution without a separate agreement. That is a looser arrangement than the custom community licenses Meta has used on past model families, and it lets teams embed the weights directly inside a shipped product.
- How does Muse Glimmer compare to Qwen3.6-27B?
- Muse Glimmer leads Qwen3.6-27B on agentic and search work — 75.5 versus 62.5 on MCP Atlas and 74.6 versus 71.1 on DeepSearch QA — and edges it on SWE-Bench Pro at 51.2 versus 50.2. Qwen3.6-27B stays ahead on SWE-Bench Verified, 77.2 to 76.0, and on MMMU Pro. The gap is widest on tool use.
- Does Muse Glimmer accept images as input?
- Yes. Muse Glimmer bundles a roughly 1.8B-parameter ViT-G/14 perception encoder alongside its 52-layer text transformer, so it takes image input as well as text. The published numbers include MMMU Pro, a multimodal reasoning benchmark, where Muse Glimmer scores 74 against 75 for Qwen3.6-27B and 73 for Gemma4-31B.
- Where can you run Muse Glimmer without hosting it yourself?
- Meta lists partner hosting for Muse Glimmer alongside the local runtimes. Together AI and Fireworks AI serve the model, while Ollama, LM Studio and vLLM cover on-device and self-hosted deployment. The weights and a Transformers loading example live on the Hugging Face model card under meta-models/Muse-Glimmer-30B.
Try it
from_pretrained("meta-models/Muse-Glimmer-30B") — or pull it in Ollama, LM Studio or vLLM