Muse Glimmer 30B
Meta · Text generation · 30B · 131k context · Released 10 August 2026
Muse Glimmer is a dense 30B model with a vision encoder, aimed at running capable agents on local hardware. It is distilled from Muse Spark, a much larger closed model behind Meta's consumer AI stack, which is how it aims to punch above its size. For a local AI audience the licence is as notable as the model: where Meta's Llama releases carry the Llama Community License and its conditions, Muse Glimmer is Apache 2.0.
Strengths
- Strong agentic behaviour, tool use, and multi-step reasoning for its size
- Runs on a single 24GB card at 4-bit, with a 128k context and image input
- Apache 2.0, so no commercial-use conditions, unlike Meta's Llama models
- Ships with a speculative-decoding drafter for faster generation
Weaknesses
- A 30B distilled model still trails the largest flagship models on the hardest tasks
- Multimodal for input only, so it reads images but does not generate them
- Benchmark and degradation figures are Meta's own, so confirm them on your own work
- New enough that community tooling and fine-tunes are still forming
Hardware requirements
| Quantisation | Approx. VRAM | Notes |
|---|---|---|
| Q4_K_M | ~18GB | Meta's 4-bit k-quant, around 17GB of weights. Meta targets a 24GB card and claims about 1% quality loss. |
| Q4 dynamic | ~26GB | Meta's higher-quality dynamic 4-bit variant. Meta targets 32GB and claims about 0.2% loss. |
| BF16 | ~62GB | Full precision. Meta cites a 64GB target. |
Also runs on CPU (slower). Optimised builds available for Apple Silicon.
What you'd need to run this
Roughly what a machine to run this would need, at up to three levels of quality. Memory is the deciding factor.
Minimum to run it
Q4_K_M · ~18GB needed
One 24GB GPU
NVIDIA Tesla P40or a Mac or mini-PC with unified memory, if you prefer no discrete GPU, Mac mini M5 Pro .
32–64GB of system RAM alongside the card.
around £700–£1,100
What else 24GB runs →For good quality
Q4 dynamic · ~26GB needed
One 32GB GPU
NVIDIA GeForce RTX 5090or a Mac or mini-PC with unified memory, if you prefer no discrete GPU, Mac mini M4 Pro .
64GB of system RAM alongside the card.
Best quality
BF16 · ~62GB needed
96GB of unified memory
AMD Ryzen AI Max+ 395 (Strix Halo)or an 80GB-class data-centre card, which is usually rented by the hour, NVIDIA A100 80GB .
Unified memory is shared with the model, so it is already counted above.
Licence
Apache 2.0 — read the licence
Benchmarks
| Benchmark | Score | Source | As of |
|---|---|---|---|
| AIME 2026 | 94.7% | Meta model card | August 2026 |
| SWE-bench Verified | 76.0% | Meta model card | August 2026 |
| GPQA Diamond | 83.5% | Meta model card | August 2026 |
How it compares
How this model’s reported scores sit against other models we cover, on the same benchmarks. This model is highlighted.
SWE-bench Verified
higher is better- Muse Glimmer 30B 76.0%
Meta model card · August 2026
- gpt-oss-120b 62.4%
OpenAI gpt-oss model card (high reasoning) · August 2025
- gpt-oss-20b 60.7%
OpenAI gpt-oss model card (high reasoning) · August 2025
- Devstral Small 53.6%
Mistral AI model card (Devstral Small 1.1) · July 2025
- Qwen3-Coder 30B-A3B 51.6
Qwen (official repo, OpenHands scaffold, 100 turns) · August 2025
- Nemotron 3.5 Lightning 51.56
NVIDIA (BF16) · August 2026
GPQA Diamond
higher is better- Qwen3.8-27B 89.2
Qwen (model card) · August 2026
- Muse Glimmer 30B 83.5%
Meta model card · August 2026
- gpt-oss-120b 80.1
OpenAI gpt-oss model card (high reasoning, no tools) · August 2025
- Nemotron 3.5 Lightning 75.44
NVIDIA (BF16) · August 2026
- DeepSeek-R1 71.5
DeepSeek-R1 technical report · January 2025
- gpt-oss-20b 71.5
OpenAI gpt-oss model card (high reasoning, no tools) · August 2025
- Qwen3 32B 68.4
Qwen3 technical report (thinking mode) · May 2025
- Qwen3 30B-A3B 65.8
Qwen3 technical report (thinking mode) · May 2025
- QwQ 32B 65.6
Qwen3 technical report (Table 13, QwQ-32B baseline) · May 2025
- Qwen3 14B 64.0
Qwen3 technical report (thinking mode) · May 2025
-
DeepSeek-R1 technical report (Table 5) · January 2025
- Qwen3 8B 62.0
Qwen3 technical report (thinking mode) · May 2025
- Gemma 3 27B 42.4
Gemma 3 technical report (27B IT) · March 2025
Muse Glimmer 30B: common questions
- What hardware do I need to run Muse Glimmer 30B?
- At its most compressed (Q4_K_M) it needs roughly 18GB of VRAM, and about 26GB for good quality. VRAM figures are approximate and depend on context length and settings.
- Is Muse Glimmer 30B free for commercial use?
- Yes. Muse Glimmer 30B is licensed under Apache 2.0, which permits commercial use with no meaningful conditions.
- Can I run Muse Glimmer 30B on Apple Silicon?
- Yes. Muse Glimmer 30B has builds optimised for Apple Silicon, through MLX or GGUF on a Mac.
- Does Muse Glimmer 30B run on CPU?
- Yes, Muse Glimmer 30B can run on the CPU, though generation is slower than on a GPU.
- What is Muse Glimmer 30B's context window?
- Muse Glimmer 30B has a context window of 131,072 tokens, about 131k.
Availability
Recommended for
- Local agentic workflows and tool use on a 24GB card
- A permissively licensed Meta model for commercial work
- Tasks that combine image understanding with agentic behaviour, run locally
Related models
Related guides
Glossary
Catalogue entry last verified 10 August 2026. Specifications change; verify anything you are about to spend money on.