Muse Glimmer 30B is the best local writer in English, but only in its larger build
Meta's open model is the top local model for English writing at 77.4 points, ahead of every Qwen and Gemma build we run. The catch: only the K-Quant Dynamic build gets there. The smaller 17GB build scores 63.8, wins the direct duels in French, German and Spanish, and is the one to download everywhere except in English.
Verdict first. Muse Glimmer 30B is a capable local generalist that produces serviceable prose on hardware you own, with no cloud round trip. In August, with 100 judgments, we called it interesting. With 1,500 ranked duels per build and language it is now the best local model in English and among the top five in the other three languages. It is still not a frontier model: the best cloud models score 94 to 99 in every language, Muse Glimmer 56 to 77.
Which file to download: Dynamic wins English by 14 points, 17GB wins the other three languages
The two builds on the main ranking
Score of each build against the whole field of the edition; rank in small print.
- K-Quant 17GB
- K-Quant Dynamic
View as a data table
| K-Quant 17GB | K-Quant Dynamic | |
|---|---|---|
| English | 63.8 (28th) | 77.4 (18th) |
| French | 72.6 (25th) | 68.5 (30th) |
| German | 65.9 (32nd) | 61.3 (37th) |
| Spanish | 62.2 (32nd) | 56.4 (41st) |
Meta ships two official text quantizations. On the English leaderboard, where each build has met 55 other local and cloud models, K-Quant Dynamic scores 77.4 and ranks 18th; K-Quant 17GB scores 63.8 and ranks 28th. In French, German and Spanish the order flips and 17GB leads by 4 to 6 points.
The direct duels say the same: Dynamic takes 63% of the points in English and 35 to 40% elsewhere
Dynamic against 17GB in direct duels
Outcome of the 26 direct duels per edition, seen from the Dynamic build.
- Dynamic wins
- Ties
- 17GB wins
View as a data table
| Dynamic wins | Ties | 17GB wins | Dynamic share | |
|---|---|---|---|---|
| English | 14 | 5 | 7 | 63.5 % |
| French | 10 | 0 | 16 | 38.5 % |
| German | 10 | 1 | 15 | 40.4 % |
| Spanish | 8 | 2 | 16 | 34.6 % |
The two builds have met each other 26 times per language. In English, Dynamic wins 14, ties 5 and loses 7. In the other three languages it loses 15 or 16 of 26. Two independent methods, the leaderboard and the head-to-head, agree in every language. We have no explanation to offer: Meta reports lower average degradation for Dynamic, and in three languages out of four our judges prefer the file with the higher reported degradation.
Against the local rivals: beats Nemotron, Qwen 3.6 and Gemma 4 31B, splits with Qwen3.8
Muse Glimmer against five local rivals
Share of points each Muse Glimmer build takes in direct duels with the rival (a win counts 1, a tie ½); about 20 duels per bar.
- K-Quant 17GB
- K-Quant Dynamic
View as a data table
| K-Quant 17GB | K-Quant Dynamic | |
|---|---|---|
| Gemma 4 31B Q8_0 | 64.3 % (21) | 77.5 % (20) |
| Qwen3.6 27B Q4_K_M | 65.0 % (20) | 71.1 % (19) |
| Qwen3.6 35B A3B Q4_K_M | 57.1 % (21) | 81.0 % (21) |
| Qwen3.8 27B Q4_K_M | 42.9 % (14) | 57.1 % (14) |
| Nemotron 3 Nano 30B A3B Q4_K_M | 95.0 % (20) | 100.0 % (20) |
In English, Muse Glimmer beats Nemotron 3 Nano almost every time, takes 57 to 81% of the points against the two Qwen 3.6 builds, and 64 to 78% against Gemma 4 31B Q8_0. Qwen3.8 27B is the newer, tougher rival: the 17GB build loses to it, the Dynamic build edges it, on 14 duels each. In German and Spanish, Gemma 4 31B turns the tables and beats both Muse builds, which matches those leaderboards, where Gemma 4 31B is the best or second-best local model.
What it is, what it costs, and whether to use it for writing
Meta Superintelligence Labs released Muse Glimmer 30B on 10 August 2026 as an open agentic model; Meta makes no prose-quality claim for it. It is a dense transformer of about 29.6 billion parameters, 1.8 billion of them in a separate vision encoder, with a 131,072-token context and a knowledge cut-off of 4 January 2026. Meta says the training data spans more than 100 languages but publishes no list, which our results by language make relevant.
The 17GB build is a 16.8 GB file (18.2 GB with the vision encoder) and needs about 20 GB with a full 128K context; Dynamic is 19.7 GB (21.1 GB) and needs about 23 GB. Text-only writers can skip the optional 1.4 GB perception encoder and the 1.6 GB speculative drafter. Both builds are slow on our Strix Halo machine, 8 to 9 tokens per second with reasoning tokens included, where Gemma 4 26B A4B produces 20 to 40, at a lower English score (44 to 57).
If you write in English and have 32 GB, Dynamic is the best local writer we have measured. If you have 24 GB, or write mostly in French, German or Spanish, take 17GB, and expect Gemma 4 31B or Qwen3.8 27B to match or beat it in those languages.
Conclusion
Muse Glimmer 30B is a serious local writer with a quirk: its two official builds do not rank in the same order across languages. Dynamic wins English by 14 points and is the top local model there; 17GB wins French, German and Spanish. With 24 GB, take 17GB; with 32 GB, take Dynamic only if you write mostly in English. It is not a frontier model: the best cloud models score 94 to 99, Muse Glimmer 56 to 77.
