Muse Glimmer 30B
Muse Glimmer 30B is **Meta's first open _agentic_ model**, released 2026-08-10 under the **Apache-2.0** license — a genuinely permissive, commercial-friendly license that's unusual for a frontier-lab release (Llama's own community license has use restrictions; this one does not). Source: [Meta AI research blog](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model) (as of 2026-08-10).
Verdict: a strong day-one open release — the best general (non-code-tuned) model on our RTX 5090 coding board.
On our first-party HumanEval+ run it scored 91.5% pass@1 (97.6% on base HumanEval), which places it 3rd of 16 on our RTX 5090 leaderboard — behind only Gemma 4 31B and Gemma 4 26B-A4B (both 93.3%) and ahead of every dedicated coding model we've tested, including Qwen3-Coder-30B (87.8%). That's notable because Muse Glimmer is a general agentic model, not a code specialist.
Throughput was a steady 80.1 tok/s at Q4_K_M on the 5090, with a 16.7 GB VRAM footprint — so it fits comfortably on a single 24 GB card.
Honest caveats: (1) day-one tooling is rough — the Ollama registry still 404s, the official Hugging Face repo is gated, and vLLM 0.26 doesn't register the architecture yet; the only clean path is llama.cpp built from master plus Meta's ungated GGUF. (2) It's a reasoning model, so tokens-per-answer (and therefore latency/cost) run higher than a non-thinking model of the same size. (3) We tested Meta's DFlash speculative-decoding drafter and saw no measurable speedup over the baseline in the day-one llama.cpp build — spec-decode integration for this arch is still maturing.
Bottom line: if you want one open, Apache-licensed, multimodal model that reasons well and codes at near-Gemma-4 level on a single consumer GPU, Muse Glimmer is an easy recommendation — just budget for the reasoning tokens.
Overview
Muse Glimmer 30B is **Meta's first open _agentic_ model**, released 2026-08-10 under the **Apache-2.0** license — a genuinely permissive, commercial-friendly license that's unusual for a frontier-lab release (Llama's own community license has use restrictions; this one does not). Source: [Meta AI research blog](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model) (as of 2026-08-10). It's a **30B dense** decoder with native **multimodality** (text + vision) and a **131,072-token** context window. Architecturally it is a new family — `MuseGlimmerForConditionalGeneration` (`model_type: muse_glimmer`): 52 layers, 32 attention heads over just **2 KV heads** (aggressive grouped-query attention), 208 head-dim, and a large 202,048-token vocabulary. Like Qwen3 and DeepSeek-R1-class models, Muse Glimmer is a **reasoning model**: it emits an internal reasoning pass before its final answer. In practice that means you must give it a generous output-token budget — at a tight `max_tokens` (< ~2K) the visible answer can come back empty because the whole budget is spent on reasoning. We benchmarked it **day-one on an NVIDIA RTX 5090** with the same HumanEval+ methodology we use across our leaderboard.
How to run it
Day-one path (the only clean one as of 2026-08-10):
The Ollama registry 404s and the official HF repo (meta-llama/muse-glimmer) is gated, so use llama.cpp from master (which already ships src/models/muse-glimmer.cpp) plus Meta's ungated GGUF mirror:
# 1. Build llama.cpp (commit 4dee52f or newer — has muse-glimmer support)
git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
cmake -B build -DGGML_CUDA=ON && cmake --build build -j --config Release
# 2. Pull the ungated Q4_K_M GGUF (~17GB) — no HF token needed
huggingface-cli download meta-models/Muse-Glimmer-30B-GGUF \
muse-glimmer-30B-kquant-17gb.gguf --local-dir ./mg
# 3. Serve (give it a big context; it's a reasoning model)
./build/bin/llama-server -m ./mg/muse-glimmer-30B-kquant-17gb.gguf \
-c 16384 -ngl 99 --host 127.0.0.1 --port 9099
Then hit http://127.0.0.1:9099/v1/chat/completions with a large max_tokens (≥ 8192) — the reasoning pass consumes budget before the answer appears.
Hardware guidance
Fits a single consumer GPU. At Q4_K_M the weights load in 16.7 GB VRAM (measured on an RTX 5090), leaving headroom for context on any 24 GB card (RTX 4090 / 5090 / 3090). For the full 131K context or the Q8_0 quant, plan for 32 GB+. Measured throughput at Q4 on the RTX 5090 was ~80 tok/s — comfortable for interactive agentic use, though the reasoning pass means each answer emits more tokens than a same-size non-thinking model.
Runtime recommendation
Run it on llama.cpp built from master (Ollama/vLLM don't support the arch day-one). Q4_K_M is the sweet spot on 24 GB cards. Note: Meta's DFlash speculative-decode drafter did not produce a measurable speedup in our day-one llama.cpp test — skip it until the integration matures.
Strengths
- Apache-2.0 — fully commercial, no gated license restrictions
- Best non-code-tuned model on our HumanEval+ board (91.5%, rank 3 of 16)
- Native vision + 131K context in a 30B dense package
- Reasoning model: strong step-by-step problem solving
- Fits a single 24GB GPU at Q4 (16.7GB VRAM measured)
Weaknesses
- Reasoning tokens inflate latency and cost per answer; needs a large max_tokens budget (empty output under ~2K tokens)
- Day-one tooling rough: Ollama 404s, official HF repo gated, vLLM 0.26 lacks the arch
- DFlash speculative-decode drafter gave no measurable speedup in day-one llama.cpp
- Slightly behind Gemma 4 31B on HumanEval+ (91.5 vs 93.3)
Reviewed quality benchmarks
First-party rows were run by RunLocalAI; reviewed community rows are labeled in the data. Every row links to the raw test-run log.
| Benchmark | Quant | Runtime / Hardware | Score | Raw log |
|---|---|---|---|---|
HumanEval tested 2026-08-10 | Q4_K_M | llama.cpp-4dee52f-vast5090 rtx-5090 | 97.6/100 | Gist → |
HumanEval+ tested 2026-08-10 | Q4_K_M | llama.cpp-4dee52f-vast5090 rtx-5090 | 91.5/100 | Gist → |
Q4_K_M note:Base HumanEval pass@1 (evalplus). Companion to the HumanEval+ run above.
Q4_K_M note:Day-one first-party run. Muse Glimmer is a reasoning model, so generation used `--max-tokens 16384` (empty answers below ~2K). Ranks 3rd of 16 on our RTX 5090 board; best non-code-tuned model tested.
Want to verify? Every row links to its Gist with full stdout and stderr of the run. The runner script is in the public repo (scripts/run-humaneval-plus.ts) — reproducible end-to-end. Browse all coding scores at /benchmarks/coding.
Quantization variants
Each quantization trades model quality for file size and VRAM. Q4_K_M is the most popular starting point.
| Quantization | File size | VRAM required |
|---|---|---|
| Q4_K_M | 17.0 GB | 20 GB |
| Q8_0 | 32.0 GB | 36 GB |
Get the model
HuggingFace
Original weights
Source repository — direct quantization required.
Hardware that runs this
Cards with enough VRAM for at least one quantization of Muse Glimmer 30B.
Models worth comparing
Same parameter band, plus what's one tier above and below — so you can decide what actually fits your hardware.
Frequently asked
What's the minimum VRAM to run Muse Glimmer 30B?
Can I use Muse Glimmer 30B commercially?
What's the context length of Muse Glimmer 30B?
Does Muse Glimmer 30B support images?
Source: huggingface.co/meta-models/Muse-Glimmer-30B
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify model claims.
Related — keep moving
Verify Muse Glimmer 30B runs on your specific hardware before committing money.