Muse Glimmer 30B writing test: how it compares to local rivals
Updated with 1,500 ranked duels per build and language. The K-Quant Dynamic build is the best local model for English writing at 77.6 points; in French, German and Spanish the smaller 17GB build is the one to download.
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 78.
The head to head: K-Quant 17GB versus Dynamic
Meta ships two official text quantizations, so the first question is which file to download. Two kinds of evidence answer it. The direct duels between the two builds, 26 per language, and the two builds’ scores on the main rankings, where each build has met 55 other local and cloud models.
| Edition | Dynamic’s share in direct duels | 17GB score (rank) | Dynamic score (rank) |
|---|---|---|---|
| English | 63.5 % (14 wins, 5 ties, 7 losses) | 64.4 (28th) | 77.6 (18th) |
| French | 38.5 % (10, 0, 16) | 73.3 (25th) | 69.0 (29th) |
| German | 40.4 % (10, 1, 15) | 65.9 (32nd) | 61.4 (37th) |
| Spanish | 34.6 % (8, 2, 16) | 63.0 (31st) | 56.4 (40th) |
Both methods agree, in every language. Dynamic wins English clearly; 17GB wins the other three. The August version of this article, on 25 duels per language, already showed the same split. It has not moved, and the rankings, built on sixty times more data, confirm it. 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
Point share of each Muse Glimmer build in direct duels, about 20 per cell. A win counts one point, a tie half.
| Opponent | EN 17GB / Dyn | FR | DE | ES |
|---|---|---|---|---|
| Gemma 4 31B Q8_0 | 64 / 78 | 80 / 81 | 45 / 52 | 33 / 41 |
| Qwen3.6 27B Q4_K_M | 65 / 71 | 75 / 75 | 58 / 63 | 55 / 71 |
| Qwen3.6 35B A3B Q4_K_M | 57 / 81 | 75 / 85 | 67 / 65 | 63 / 73 |
| Qwen3.8 27B Q4_K_M | 43 / 57 | 62 / 35 | 71 / 50 | 82 / 50 |
| Nemotron 3 Nano 30B A3B Q4_K_M | 95 / 100 | 95 / 100 | 100 / 91 | 91 / 90 |
Muse Glimmer beats Nemotron 3 Nano almost every time and beats Qwen 3.6 comfortably in English and French. Gemma 4 31B Q8_0 is the rival that turns the tables: Muse Glimmer wins against it in English and French and loses in German and Spanish, which matches the leaderboards, where Gemma 4 31B is the best local model in German. Qwen3.8 27B is the newer, tougher rival; the sample is small (10 to 17 duels) and the direction changes with the language.
What Muse Glimmer 30B actually is
Meta Superintelligence Labs released the model on 10 August 2026 as an open agentic model. Meta makes no prose quality claim for it. The model card describes a dense transformer paired with a separate perception encoder, approximately 29.6 billion parameters of which roughly 1.8 billion sit in the vision encoder, across 52 text layers. The configured context is 131,072 tokens; the knowledge cutoff is 4 January 2026. Meta says training data spans more than 100 languages but publishes no language list, which our own results by language make relevant.
Memory and speed
| Build | GGUF file | With a full 128K context | Meta’s VRAM target | Measured speed (Strix Halo) |
|---|---|---|---|---|
| K-Quant 17GB | 15.6 GB | 22.1 GB | 24 GB | 8.8 tok/s |
| K-Quant Dynamic | 18.3 GB | 24.8 GB | 32 GB | 8.0 tok/s |
The optional perception encoder adds 1.4 GB and is needed only for images; the optional speculative drafter adds 1.6 GB. Text-only writers can skip both. Both builds are slow on our machine, at 8 to 9 tokens per second with reasoning tokens included: Gemma 4 26B A4B produces 20 to 40 tokens per second on the same hardware at a similar English score for its Q6_K build.
Should you use it for writing?
For English, yes, and take the Dynamic build if you have 32 GB. For French, German and Spanish, the 17GB build is a good choice for a 24 GB card, competitive with Gemma 4 31B and Qwen3.8 27B, which are the alternatives worth testing on your own texts. As with every model on this site, nothing here says anything about factual accuracy; the duels judge writing, and factual claims still need checking.