Best model for writing
Models ranked by the mean of the two judges' writing scores, 0 to 10. Only ranked models that both judges scored are included, and models are grouped by total parameters.
Read the memory figure as an estimate, not a guarantee. The memory a model needs is estimated from its file: the weights plus about 20% for the context of a single conversation. It is not a measurement, and a long conversation or several at once needs more. A model with no file size on record is not counted as fitting any amount of memory. Speeds, and the GPU memory shown on the speed page, were measured on this benchmark's server, not on yours. That server kept 1 to 8 answers in flight at a time and loaded a context window of 16,384 tokens for most models; 42 of the models ranked ran at a shorter context. Speeds were measured on one laptop and are comparable only between models on the same backend. See the settings and the memory method.
All sizes
gemma-4-31B-it leads here with 7.0 on a scale of 0 to 10; its Overall is 99.3 (95% interval 98.3–100.0).
| Place | Model | Writing (0 to 10) | Overall | Rank range | Decode (tok/s) | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | gemma-4-31B-it | 7.0 | 99.398.3–100.0 | 1–5 | 3.8ROCm 7.1 | 22.4 | 16–35B |
| 1 | Muse Glimmer 30B | 7.0 | 97.996.0–99.3 | 1–23 | 3.6ROCm 7.1 | 21.8 | 16–35B |
| 3 | gemma-3-27b-it | 6.9 | 95.893.7–97.5 | 3–36 | 2.7ROCm 7.1 | 20.9 | 16–35B |
| 4 | gemma-4-26B-A4B-it | 6.8 | 97.895.8–99.3 | 1–22 | 15.7ROCm 7.1 | 21.6 | 16–35B |
| 4 | Qwen3.5-27B | 6.8 | 96.293.8–98.3 | 3–35 | 2.7ROCm 7.1 | 19.9 | 16–35B |
Tiny (under 3B)
Granite3.3 2B leads its size class with 5.2 on a scale of 0 to 10; its Overall is 68.6 (95% interval 63.8–73.8).
| Place | Model | Writing (0 to 10) | Overall | Rank range | Decode (tok/s) | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | Granite3.3 2B | 5.2 | 68.663.8–73.8 | 111–159 | 28.6ROCm 7.1 | 1.8 | Tiny (under 3B) |
| 1 | Lfm3 VL 2B | 5.2 | 72.267.2–77.0 | 89–158 | 41.9ROCm 7.1 | 1.8 | Tiny (under 3B) |
| 3 | Gemma2 2B Instruct | 5.1 | 68.262.9–73.3 | 112–160 | 27.6ROCm 7.1 | 2.0 | Tiny (under 3B) |
| 3 | Gemma3 1B IT | 5.1 | 56.851.8–62.2 | 154–188 | 57.0ROCm 7.1 | 1.0 | Tiny (under 3B) |
| 3 | Exaone3.5 2.4B Instruct | 5.1 | 76.370.9–81.3 | 73–141 | 100.9ROCm 7.1 | 1.9 | Tiny (under 3B) |
3–4B
Ministral 3 3B Instruct 2512 leads its size class with 5.7 on a scale of 0 to 10; its Overall is 80.9 (95% interval 76.5–85.2).
| Place | Model | Writing (0 to 10) | Overall | Rank range | Decode (tok/s) | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | Ministral 3 3B Instruct 2512 | 5.7 | 80.976.5–85.2 | 61–124 | 70.9ROCm 7.1 | 3.6 | 3–4B |
| 2 | Phi-3-mini-4k-instruct | 5.6 | 63.458.3–68.1 | 129–182 | 17.6ROCm 7.1 | 2.9 | 3–4B |
| 3 | Llama-3.2-3B-Instruct | 5.5 | 75.771.0–80.4 | 84–150 | 20.7ROCm 7.1 | 2.4 | 3–4B |
| 4 | SmolLM3-3B | 5.3 | 78.073.0–83.1 | 71–142 | 22.2ROCm 7.1 | — | 3–4B |
| 5 | Granite4 3B | 5.2 | 74.169.0–79.2 | 85–145 | 71.3ROCm 7.1 | 2.5 | 3–4B |
4–9B
Gemma4 E2B IT leads its size class with 6.8 on a scale of 0 to 10; its Overall is 93.8 (95% interval 91.0–96.1).
| Place | Model | Writing (0 to 10) | Overall | Rank range | Decode (tok/s) | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | Gemma4 E2B IT | 6.8 | 93.891.0–96.1 | 4–61 | 17.2ROCm 7.1 | 5.5 | 4–9B |
| 2 | Gemma3n E2B IT | 6.3 | 79.975.2–84.7 | 64–127 | 13.8ROCm 7.1 | 6.7 | 4–9B |
| 2 | Granite4.2 8B | 6.3 | 81.977.2–86.3 | 59–128 | 11.1ROCm 7.1 | 6.4 | 4–9B |
| 4 | Gemma 3n E4B IT | 6.2 | 86.182.4–89.6 | 44–97 | 12.1ROCm 7.1 | 9.1 | 4–9B |
| 4 | Nemotron 3.5 Bolt E4B | 6.2 | 90.386.3–93.9 | 8–82 | 9.8ROCm 7.1 | 8.6 | 4–9B |
9–16B
Gemma 4 12B IT leads its size class with 6.8 on a scale of 0 to 10; its Overall is 98.6 (95% interval 97.0–99.8).
| Place | Model | Writing (0 to 10) | Overall | Rank range | Decode (tok/s) | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | Gemma 4 12B IT | 6.8 | 98.697.0–99.8 | 1–12 | 8.1ROCm 7.1 | 8.9 | 9–16B |
| 2 | Qwen3.5-9B | 6.4 | 92.089.0–95.1 | 11–69 | 8.4ROCm 7.1 | 6.8 | 9–16B |
| 3 | Ministral 3 14B Instruct (2512) | 6.3 | 92.790.0–95.2 | 9–61 | 5.6ROCm 7.1 | 10.9 | 9–16B |
| 3 | Qwen3-14B | 6.3 | 91.788.5–94.7 | 11–71 | 5.0ROCm 7.1 | 11.1 | 9–16B |
| 5 | gemma-3-12b-it | 6.2 | 92.789.7–95.4 | 8–66 | 6.1ROCm 7.1 | 9.8 | 9–16B |
16–35B
gemma-4-31B-it leads its size class with 7.0 on a scale of 0 to 10; its Overall is 99.3 (95% interval 98.3–100.0).
| Place | Model | Writing (0 to 10) | Overall | Rank range | Decode (tok/s) | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | gemma-4-31B-it | 7.0 | 99.398.3–100.0 | 1–5 | 3.8ROCm 7.1 | 22.4 | 16–35B |
| 1 | Muse Glimmer 30B | 7.0 | 97.996.0–99.3 | 1–23 | 3.6ROCm 7.1 | 21.8 | 16–35B |
| 3 | gemma-3-27b-it | 6.9 | 95.893.7–97.5 | 3–36 | 2.7ROCm 7.1 | 20.9 | 16–35B |
| 4 | gemma-4-26B-A4B-it | 6.8 | 97.895.8–99.3 | 1–22 | 15.7ROCm 7.1 | 21.6 | 16–35B |
| 4 | Qwen3.5-27B | 6.8 | 96.293.8–98.3 | 3–35 | 2.7ROCm 7.1 | 19.9 | 16–35B |
> 35B
gpt-oss-120b leads its size class with 6.7 on a scale of 0 to 10; its Overall is 98.6 (95% interval 96.8–100.0).
| Place | Model | Writing (0 to 10) | Overall | Rank range | Decode (tok/s) | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | gpt-oss-120b | 6.7 | 98.696.8–100.0 | 1–14 | 11.0Vulkan | 78.4 | > 35B |
| 2 | Sage-A1.5-Q4_K_M | 6.6 | 94.591.8–96.7 | 5–56 | 22.6ROCm 7.1 | 26.4 | > 35B |
| 3 | Flint 40B A2B | 6.4 | 89.085.1–92.6 | 14–90 | 25.9ROCm 7.1 | 27.6 | > 35B |
| 4 | Llama-3.3-70B-Instruct | 6.3 | 94.391.3–96.7 | 4–63 | 1.4Vulkan | 51.0 | > 35B |
| 4 | Qwen3-Coder-Nextcode model | 6.3 | 96.494.2–98.3 | 2–42 | 18.1Vulkan | 62.1 | > 35B |
A higher place is not always a clear win. When two models have overlapping rank ranges, this battery cannot tell them apart; see how sure a score is. A dash means not measured or not scored, which is different from 0. Speeds were measured on one laptop and are comparable only between models on the same backend.
More: all the best-model pages, the hardware picker, speed and memory for every model.