Fastest model that is still good
Ranked models whose Overall score is within 10 points of the best ranked model, fastest first. Speed is single-stream decode speed in tokens per second, measured on one laptop.
The best ranked model scores 99.3, so the window is Overall 89.3 to 99.3. 55 of 227 ranked models are inside it. 1 of them has no single-stream speed measurement and is not listed.
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.
Fastest within the quality window
Qwen3-VL-30B-A3B-Instruct is the fastest of the 54 models in the window: 26.9 tokens per second (measured on Vulkan; speeds on different backends are not directly comparable), with an Overall of 92.9 (95% interval 90.1–95.5), 6.4 points below the best ranked model, gemma-4-31B-it (99.3).
| Place | Model | Decode (tok/s) | Overall | Rank range | Points below the best | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | Qwen3-VL-30B-A3B-Instruct | 26.9Vulkan | 92.990.1–95.5 | 7–65 | 6.4 | 23.5 | 16–35B |
| 2 | Qwen3.5-35B-A3B | 23.2Vulkan | 94.892.1–97.1 | 3–60 | 4.5 | 25.4 | 16–35B |
| 3 | Laguna XS 2.1 | 23.1ROCm 7.1 | 91.287.7–94.4 | 6–73 | 8.1 | 24.3 | — |
| 4 | Sage-A1.5-Q4_K_M | 22.6ROCm 7.1 | 94.591.8–96.7 | 5–56 | 4.8 | 26.4 | > 35B |
| 5 | Qwen3.6-35B-A3B | 22.1Vulkan | 94.691.6–97.2 | 3–63 | 4.7 | 28.7 | 16–35B |
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.