Overview
Muse Glimmer 30B is an agent model from Meta Superintelligence Lab, distilled from Muse Spark and released in August 2026 under Apache 2.0. Meta built it for autonomous agent work on consumer hardware, so the design target was a machine you own rather than a cluster you rent.
| Specification | Muse Glimmer 30B |
|---|---|
| Total parameters | 29.6B, including the vision encoder |
| Architecture | Dense causal transformer with a perception encoder |
| Layers | 52 |
| Attention | Sliding window of 2,048 tokens on three of every four layers |
| Perception encoder | ViT-G/14, about 1.8B parameters |
| Context window | 131,072 tokens |
| Modalities | Text and image input, text output |
| Speculative decoding | DFlash drafter ships with the model |
| Release date | August 2026 |
| License | Apache 2.0 |
That attention layout is what keeps the KV cache small enough for the vision encoder and the DFlash drafter to sit alongside the weights in a 24 GB envelope. Meta reports that its own 4-bit compression costs 0.2% accuracy against full precision on a 15-benchmark average.
Muse Glimmer 30B benchmarks
Meta's launch numbers, from the model card, put Glimmer in high-reasoning mode against the thinking modes of Gemma 4 31B and Qwen3.6-27B:
| Muse Glimmer 30B | Gemma 4 31B | Qwen3.6-27B | |
|---|---|---|---|
MCP Atlas Tool use over MCP servers | 75.5 | 54.2 | 62.5 |
DeepSearch QA Multi-step web research | 74.6 | 61.7 | 71.1 |
SWE-Bench Pro Real-world software engineering, harder split | 51.2 | 36.9 | 50.2 |
SWE-Bench Verified Real-world software engineering tasks | 76.0 | 66.6 | 77.2 |
TerminalBench 2.1 Agentic terminal-use tasks | 51.7 | 43.4 | 60.7 |
AIME 2026 Advanced mathematical reasoning | 94.7 | 89.2 | 94.1 |
GPQA Diamond Expert-level science questions | 83.5 | 85.7 | 84.2 |
OSWorld-Verified Computer-use agent tasks | 65.9 | 58.5 | 75.6 |
Glimmer takes tool use, web research and maths, and gives up terminal and computer-use tasks to Qwen3.6-27B. The pitch is a 30B agent that stays inside a 24 GB card with its vision encoder and drafter loaded, not a model that wins every row.
Muse Glimmer 30B hardware requirements
The system requirement to check is memory. We quantized the model from the original weights and published the builds as AtomicChat/Muse-Glimmer-30B-GGUF.
| Memory | Build to pick | File size |
|---|---|---|
| 16 GB | AD-IQ3_M | 14.9 GB |
| 24 GB | AD-Q4_K_M | 19.1 GB |
| 32 GB | AD-Q5_K_M | 22.7 GB |
| 48 GB and up | AD-Q6_K | 28.3 GB |
Add the vision projector on top if you want image input, and leave headroom for the DFlash drafter if you run speculative decoding. Neighbouring files differ by a gigabyte or two, so when two builds both fit, take the larger one.
How to run Muse Glimmer 30B in Atomic Chat
Atomic Chat is a free local app for macOS, Windows and Linux. It includes a Hugging Face model browser and a built-in chat, with no manual llama.cpp build required.
- Download Atomic Chat for your platform and open it.
- Search for Muse-Glimmer-30B in the model browser and open Download Options.
- Pick the build that fits the memory you have, then start a chat.
To point it at real tools once it is running, follow our guide to running AI agents locally, or connect it to your apps with MCP connectors.
Its closest rival at this size is Ornith-1.5-35B-A3B, which is ahead on coding and behind on MCP tool routing. For the wider field see the best local LLMs for coding.
License
Muse Glimmer 30B is released under Apache 2.0, which permits commercial use, modification and redistribution with no royalties. That is unusually permissive for a Meta model release and it applies to the perception encoder and the DFlash drafter shipped with it.