What people run, by hardware
Pick the kind of machine you have. You'll see the model sizes people report running on it, most reported first, with the speed they typically get. It tells you what works on hardware like yours. It doesn't tell you which model gives the best answers.
To check a specific model on your exact card, use the calculator. For coding, see the coding models by VRAM.
AMD and Intel 20-32 GB· AMD only
Discrete AMD and Intel cards from 20 to 32 GB. Memory speed 512 to 1,024 GB/s across the 5 devices with reports, so a typical figure here spans all of them. 76 single-device reports.
| # | Model | Reports | Gen t/s | PP t/s | With speculation |
|---|---|---|---|---|---|
| 1 | Qwen3.827B Alibaba · mostly Q4_K_M Gen 34 t/s (13 runs) · PP 814 t/s (6 runs) · with speculation 69 t/s (12) | 34 | 34 13 runs, 3 devices | 814 6 runs | 69 12 runs |
| 2 | Qwen3.8 Flash-Next125B · 6B active Alibaba · mostly 3.05bpw No generation data | 10 | no generation data | no PP data | none |
| 3 | Qwen3.8 Swift-1.527B Alibaba · mostly MXFP4 Gen 38 t/s (1 run) · with speculation 112 t/s (3) | 5 | 38 1 run | no PP data | 112 3 runs |
| 4 | Gemma 426B · 4B active Google DeepMind · mostly Q4_0 Gen 143 t/s (1 run) · PP 3766 t/s (2 runs) | 3 | 143 1 run | 3766 2 runs | none |
| 5 | Qwen3.635B · 3B active Alibaba · mostly INT4 Gen 84 t/s (3 runs) · PP 2000 t/s (3 runs) | 3 | 84 3 runs, 2 devices | 2000 3 runs | none |
| 6 | Ornith1.535B · 3B active mostly MXFP4 Gen 109 t/s (1 run) · with speculation 245 t/s (1) | 2 | 109 1 run | no PP data | 245 1 run |
| 7 | Gemma 4 E4B8B Google DeepMind · mostly FP16 Gen 58 t/s (1 run) · PP 83 t/s (1 run) | 2 | 58 1 run | 83 1 run | none |
| 8 | Qwen3.535B · 3B active Alibaba · mostly UD-Q4_K_XL Gen 149 t/s (1 run) · PP 3075 t/s (1 run) | 1 | 149 1 run | 3075 1 run | none |
| 9 | Qwen2.57B Alibaba · mostly GPTQ-Int4 Gen 115 t/s (1 run) | 1 | 115 1 run | no PP data | none |
| 10 | Qwen3-Coder30B · 3B active mostly Q4_K_M Gen 66 t/s (1 run) | 1 | 66 1 run | no PP data | none |
| 11 | Qwen3 Thinking-250730B · 3B active Alibaba · mostly Q4_0 Gen 63 t/s (1 run) · PP 1224 t/s (1 run) | 1 | 63 1 run | 1224 1 run | none |
| 12 | Qwen2.51.5B Alibaba · mostly FP16 Gen 32 t/s (1 run) | 1 | 32 1 run | no PP data | none |
| 13 | Gemma 431B Google DeepMind · mostly Q4_K_XL Gen 30 t/s (1 run) · PP 422 t/s (1 run) | 1 | 30 1 run | 422 1 run | none |
| 14 | Qwen3.8 Swift27B Alibaba · mostly Q4_K_XL Gen 25 t/s (1 run) · PP 492 t/s (1 run) | 1 | 25 1 run | 492 1 run | none |
| 15 | Raon-1 mini-preview29B · 4B active mostly MQ4 Gen 4.6 t/s (1 run) · PP 13 t/s (1 run) | 1 | 4.6 1 run | 13 1 run | none |
| 16 | DeepSeek V4.1 Flash552B · 16B active mostly IQ1_M/IQ2_XS/IQ2_XXS (IST-DAS 2.00bpw) No generation data | 1 | no generation data | no PP data | none |
| 17 | GLM-5.3 Flash320B · 18B active Zhipu AI · mostly Maya-S No generation data | 1 | no generation data | no PP data | none |
| 18 | Gemma 426B Google DeepMind No generation data | 1 | no generation data | no PP data | none |
| 19 | Ornith1.5122B mostly UD-IQ2_XXS No generation data | 1 | no generation data | no PP data | none |
| 20 | Qwen3.627B Alibaba · mostly Q6 No generation data · PP 32 t/s (1 run) | 1 | no generation data | 32 1 run | none |
| 21 | Qwen3.6 Fable-Fus-711-UnHeretic-NM-DAU-NEO-MAX-NEO-LOW27B Alibaba · mostly IQ4_XS No generation data · with speculation 42 t/s (1) | 1 | no generation data | no PP data | 42 1 run |
| 22 | Qwen3.8 CODER27B Alibaba · mostly IQ4_XS No generation data | 1 | no generation data | no PP data | none |
| 23 | Qwen3.8 Uncensored-HauhauCS-Aggressive27B Alibaba · mostly Q5_K_P No generation data · with speculation 14 t/s (1) | 1 | no generation data | no PP data | 14 1 run |
| 24 | Qwen3.8 Uncensored27B Alibaba No generation data | 1 | no generation data | no PP data | none |
Not in the table: 8 reports don't say which model size ran. 65 ran on more than one device, which says nothing about what one of them does.
How this list works
Each row is one model size, because a 27B and a 397B from the same family are different choices. Rows are ordered by how many people report running that size on this hardware.
Generation speed is the median of runs on one device, serving one request, without speculative decoding. PP is the median prompt-processing speed from plain runs that reported it, so its run count can be smaller. Speculative decoding (MTP, a draft model, DFlash, n-gram lookup) can make the same model several times faster, so those runs get their own column. A report that doesn't make clear which kind of run it was counts in neither.
Hardware is grouped by how much it holds and how fast its memory is, since memory speed sets how fast a model writes. Click a model for every report behind its row.