whatmodel.sh / Muse Glimmer 30B

Muse Glimmer 30B

+ contribute

Muse Glimmer 30B is an open-weights language model from Meta, released August 2026. Its published size is 29.8B parameters and a 131k-token context window. Licence: apache-2.0.

69.5±1.1#89 of 719 LLMsfrom 21 benchmark results across 6 sources

$ how it compares

The LLMs ranked either side of Muse Glimmer 30B, out of 828 with a score, then more from Meta.

#	model	score	where	RAM	released
95	MiMo-V2.5-Pro (1T-A42B) Xiaomi	70±1	api $0.65/M +local	~163GB	2026-04
96	Inkling (975B-A41B) Thinking Machines	70±1	api $1.73/M +local	~156GB	2026-07
97	Muse Glimmer 30B Meta	70±1	api $0.64/M +local	~19GB	2026-08
98	Nemotron 3 Ultra 550B-A55B NVIDIA	69±1	api $1.35/M +local	~89GB	2026-06
99	Qwen3.6 27B Alibaba	69±1	api $1.20/M +local	~18GB	2026-04
	more from Meta
24	Muse Spark 1.3 Meta	77±1	api $2.00/M		2026-09
40	Muse Spark 1.2 Meta	75±1	api $2.00/M		2026-08
43	Muse Spark 1.1 Meta	75±1	api $2.00/M		2026-07

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #69 of 300 · 32% of its score · 6 tests

Beats Gemini 3.6 Flash, GPT-5 Pro, LongCat 2.0 (1.6T-A48B) here, though they rank higher overall. Trails MiMo-V2.5 (310B-A15B) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
SWE-bench Pro (split unspecified) lab	51.2% ✎	▃·▁▁▁▂▂▃█▇▃▁ beats 28/55	very hard · avg 8.3%	▇▇▇▇ 9.1%	sharp · at its level · small run
SWE-bench Verified lab	76.0% ✎	▁▁▁▁▁▁▂▄█▆▁▁ beats 105/131	hard · avg 23%	▇▇ 6.1%	sharp · at its level · small run
SWE-bench Pro public set, Qwen-refined	51.2% ✎	▂·····▂·▂█▃▅ beats 5/12	hard · avg 19%	▇▇ 5.4%	at its level
OSWorld-Verified lab	65.9% ✎	█▄·▂▆▂▄▂▆▄▆█ beats 12/25	hard · avg 26%	▇▇ 4.4%	at its level · small run
MCP-Atlas (public set) lab	75.5% ✎	▂·▁▂▁·▁▃█▆▆▂ beats 30/37	hard · avg 28%	▇ 3.5%	at its level · small run
DeepSearchQA lab	74.6% ✎	█·······█··█ beats 2/2	medium · avg 40%	▇ 3.1%	at its level · small run

coding▁▁▁▃▆██▇▆▄▃▁ #105 of 458 · 48% of its score · 5 tests

Beats Grok 4.3 here, though they rank higher overall. Trails Llama 3.1 Nemotron Ultra 253B v1, Solar Pro 4, Nemotron Cascade 2 30B-A3B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena WebDev (Code Arena)	1354.8 ✎	▁▂▄▃▅█▄▇▆▄▁▁ beats 28/119	avg 981	▇▇▇▇▇▇▇ 18%	sharp
SWE-bench Pro (split unspecified) lab	51.2% ✎	▃·▁▁▁▂▂▃█▇▃▁ beats 28/55	very hard · avg 8.3%	▇▇▇▇ 9.1%	sharp · at its level · small run
Luau Leetcode fully solved (Torpedo Software)	68.3% ✎	▁▁▁▁·▁▂▃▅▃█▇ beats 19/49	hard · avg 18%	▇▇▇ 8.7%	at its level
SWE-bench Verified lab	76.0% ✎	▁▁▁▁▁▁▂▄█▆▁▁ beats 105/131	hard · avg 23%	▇▇ 6.1%	sharp · at its level · small run
SWE-bench Pro public set, Qwen-refined	51.2% ✎	▂·····▂·▂█▃▅ beats 5/12	hard · avg 19%	▇▇ 5.4%	at its level

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #82 of 393 * · 10% of its score · 1 test

Beats LongCat 2.0 (1.6T-A48B), Inkling Small (266B-A12B) here, though they rank higher overall. Trails Gemma 4 12B, Gemma 4 26B-A4B, Mistral Medium 3.5 (128B) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LMArena (text)	1425.0 ✎	▁▁▁▂▃▃▃▆▆▇█▃ beats 277/379	avg 1268	▇▇▇▇ 10%	blunt

mathnot ranked on this task yet · 1 test

test	result	vs the others in its run	difficulty	counts	why
AIME 2026 lab	94.7% ✎	▁▁▁▂▁▂▆█▆▂·▁ beats 37/49	medium · avg 48%	—	not counted yet: no other model in the same run to compare it with

vision & documents▁▁▃▂▁▃▄▂▆▇█▄ #62 of 147 · 8.6% of its score · 3 tests

Beats DeepSeek V4 Flash Vision (Exp), GPT-5, Inkling Small (266B-A12B) here, though they rank higher overall. Trails Gemma 4 26B-A4B, Claude Haiku 4.5, Qwen3.6 Flash here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
CharXiv (reasoning questions) lab	78.8% ✎	▂▅▃▅▅▆▆▅▆█▆▂ beats 74/100	medium · avg 42%	▇ 3.6%	small run
OmniDocBench 1.5 overall (higher is better) lab	75.8% ✎	▂··▂▃▂··▂▂█▅ beats 2/13	medium · avg 54%	▇ 3.2%	small run
ScreenSpot-Pro lab	75.4% ✎	▂▃▅▆██▅▂···▂ beats 20/24	medium · avg 51%	▇ 1.8%	blunt · at its level · small run

web search▂▁▂▆▄█▃▃▁▁·▁ #46 of 80 * · 3.1% of its score · 1 test

Beats Gemini 3.1 Pro (Preview), Qwen3.7 Max, GPT-5.2 here, though they rank higher overall. Trails GPT-4o mini (2024-07-18), Mistral Small 3.1 24B (2503), Llama 4 Scout (109B-A17B) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
DeepSearchQA lab	74.6% ✎	█·······█··█ beats 2/2	medium · avg 40%	▇ 3.1%	at its level · small run

other tests9 not in a task · 22% of its score

test	result	vs the others in its run	difficulty	counts	why
WildClawBench lab	47.6% ✎	▂·▄▄▆▂▄█▆▂▂▂ beats 15/20	very hard · avg 14%	▇▇ 4.2%	at its level · small run
CPTUB average (SpeakLeash, 0-5 judge scale)	88.3% ✎	▁▂▂▂▃█▃▆▅▄▂▃ beats 95/95	easy · avg 66%	▇ 3.5%	typical
RobloxQA v2.0 multiple choice (Torpedo Software)	89.9% ✎	▁▁▃▁█▄▃▃▆▂▂▂ beats 14/49	easy · avg 78%	▇ 3.2%	typical
Gaia2 lab	43.3% ✎	█·····█····█ beats 2/2	hard · avg 16%	▇ 2.8%	at its level · small run
IFBench (aggregation unspecified) lab	77.0% ✎	▃▇▂▄▅▅▂▄▄▆█▇ beats 70/80	medium · avg 40%	▇ 2.7%	at its level · small run
SkillsBench (with skills) lab	44.3% ✎	█·········██ beats 1/2	hard · avg 16%	▇ 2.7%	at its level · small run
WMT25 Terminology Translation (term success rate)	90.1% ✎	▂·▂··▂·▄▂·█▆ beats 8/12	medium · avg 64%	▇ 1.5%	noisy
WMT24++	77.7% ✎	█·▃·▅▅▃···▃▃ beats 3/10	easy · avg 71%	▇ 0.8%	blunt

1 more test

test	result	vs the others in its run	difficulty	counts	why
WMT26 General MT (GEMBA-ESA, LLM judge)	78.2% ✎	▄·▄▄▄·▄██▄█▄ beats 7/12	easy · avg 65%	▇ 0.7%	blunt

$ where to run it

Through an API

api · Fireworks AI · $0.64/M · $0.35 in / $1.5 out per 1M tokens · 131k context · fireworks.ai ↗ · price sourceStandard base token rate from models.dev. Provider-specific tiers, long-context surcharges, caching and promotions may differ.

llama.cpp · recommended

Install llama.cpp · Homebrew and winget installs name the commands llama-cli and llama-server instead.
macOS / Linux
curl -LsSf https://llama.app/install.sh | sh
Windows
irm https://llama.app/install.ps1 | iex
other ways to install ↗
Run Muse Glimmer 30B · needs ~16.8 GB · from huggingface.co/meta-models
chat in the terminal
llama cli -hf meta-models/Muse-Glimmer-30B-GGUF:Q4_K_M
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf meta-models/Muse-Glimmer-30B-GGUF:Q4_K_M

LM Studio

version

Install LM Studio · Open the app once before using the lms command.
download it (macOS / Windows / Linux) ↗
Run Muse Glimmer 30B · needs ~16.8 GB · from huggingface.co/meta-models
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/meta-models/Muse-Glimmer-30B-GGUF@Q4_K_M -y
Then run lms ls to see its name, and lms chat followed by that name

Ollama

version

Install Ollama
download it (macOS / Windows) ↗
Linux
curl -fsSL https://ollama.com/install.sh | sh
Run Muse Glimmer 30B · needs ~18.38 GB · from ollama.com
chat in the terminal
ollama run muse-glimmer:30b

MLX (Mac)

version

Install MLX (Mac) · Apple silicon Macs only, text only.
with uv
uv tool install mlx-lm
or
pip install mlx-lm
Run Muse Glimmer 30B · needs ~19.4 GB, Mac only · from huggingface.co/mlx-community
chat in the terminal
mlx_lm.chat --model mlx-community/Muse-Glimmer-30B-4bit
Or start an OpenAI-compatible server at http://localhost:8080/v1
mlx_lm.server --model mlx-community/Muse-Glimmer-30B-4bit

Weights and documentation: huggingface.co/meta-models ↗

$ how the score was worked out · meta-models/Muse-Glimmer-30B… 47%, LMArena leaderboard dataset 29%, Roblox LLM Leaderboard (Torpe… 12% and 3 more

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

source	results	share of the score	licence
meta-models/Muse-Glimmer-30B model card (speed and benchmark tables) · lab-reported	12	▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 47%	Apache-2.0
LMArena leaderboard dataset	2	▇▇▇▇▇▇▇▇▇ 29%	CC BY 4.0
Roblox LLM Leaderboard (Torpedo Software): RobloxQA v2.0 and Luau Leetcode	2	▇▇▇▇ 12%	MIT
Qwen/Qwen3.8-27B model card	1	▇▇ 5%	apache-2.0
SpeakLeash CPTUB (Complex Polish Text Understanding Benchmark)	1	▇ 3%	MIT
Cohere — North Small Translate (arXiv:2609.13916v1)	3	▇ 3%	CC BY 4.0
prior · pull toward the middle		▇ 1%
weighted average		= 69.5

not counted (1)

meta-models/Muse-Glimmer-30B model card (speed and benchmark tables)

AIME 2026 [correct]

94.7%

fewer than three models have this test yet, or this source ran it for this model alone

$ published speed measurements · 3

Reference results for the hardware, engine and settings below. They aren't verified speeds of the downloads above or predictions for your machine.

NVIDIA GeForce RTX 5090 (host CPU/RAM not stated)	K-Quant-17GB (Meta ~4-bit K-quant build of the language model; exact file/format not identified)	74.9 tok/s	llama.cpp (version not stated) · batch size 1; greedy decoding; no speculative decoding; prompt set, prompt/output lengths and context depth not stated Published single-stream decode rate without speculation. Engine versions and whether the value is a mean over the prompt set are not stated. Card: "M4/M5 measurements were done using ExecuTorch, and RTX using llama.cpp."	source ↗
Apple M4 Max (MacBook; GPU core count and unified memory not specified)	K-Quant-17GB (Meta ~4-bit K-quant build of the language model; exact file/format not identified)	23.7 tok/s	ExecuTorch (version not stated) · batch size 1; greedy decoding; no speculative decoding; prompt set, prompt/output lengths and context depth not stated Published single-stream decode rate without speculation. Engine versions and whether the value is a mean over the prompt set are not stated. Card: "M4/M5 measurements were done using ExecuTorch, and RTX using llama.cpp."	source ↗
Apple M5 Max (MacBook; GPU core count and unified memory not specified)	K-Quant-17GB (Meta ~4-bit K-quant build of the language model; exact file/format not identified)	26.6 tok/s	ExecuTorch (version not stated) · batch size 1; greedy decoding; no speculative decoding; prompt set, prompt/output lengths and context depth not stated Published single-stream decode rate without speculation. Engine versions and whether the value is a mean over the prompt set are not stated. Card: "M4/M5 measurements were done using ExecuTorch, and RTX using llama.cpp."	source ↗

$ improve this page

Update model information Add a place to run it Add a test result Propose a benchmark Correct a benchmark score Edit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated 2026-10-11 02:20 UTC · sources and licences · json