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 up to 16 GB· AMD only
Discrete AMD and Intel cards up to 16 GB, on ROCm, Vulkan or SYCL. Memory speed 224 to 1,024 GB/s across the 16 devices with reports, so a typical figure here spans all of them. 57 single-device reports.
| # | Model | Reports | Gen t/s | PP t/s | With speculation |
|---|---|---|---|---|---|
| 1 | Qwen3.827B Alibaba · mostly IQ3_XXS Gen 30 t/s (5 runs) · PP 371 t/s (2 runs) · with speculation 46 t/s (4) | 18 | 30 5 runs, 4 devices | 371 2 runs | 46 4 runs |
| 2 | Qwen3.8 Flash-Next125B · 6B active Alibaba · mostly Q2_0 No generation data | 10 | no generation data | no PP data | none |
| 3 | Qwen3.635B · 3B active Alibaba · mostly Q4_K_M No generation data · with speculation 110 t/s (1) | 8 | no generation data | no PP data | 110 1 run |
| 4 | Qwen38B Alibaba · mostly Q4_K_M Gen 80 t/s (2 runs) · PP 2548 t/s (2 runs) · with speculation 133 t/s (1) | 3 | 80 2 runs, 2 devices | 2548 2 runs | 133 1 run |
| 5 | Qwen3.59B Alibaba · mostly Q6_K_XL Gen 35 t/s (2 runs) | 3 | 35 2 runs | no PP data | none |
| 6 | Qwen3.8 Swift-1.527B Alibaba · mostly IQ3_S Gen 30 t/s (1 run) · with speculation 44 t/s (1) | 3 | 30 1 run | no PP data | 44 1 run |
| 7 | Bonsai27B PrismML · mostly Q1_0 Gen 73 t/s (1 run) | 1 | 73 1 run | no PP data | none |
| 8 | Gemma 412B Google DeepMind · mostly IQ4_NL Gen 35 t/s (1 run) · PP 654 t/s (1 run) | 1 | 35 1 run | 654 1 run | none |
| 9 | Qwen2.5 Coder14B Alibaba · mostly Q4_K_M Gen 34 t/s (1 run) · PP 1210 t/s (1 run) | 1 | 34 1 run | 1210 1 run | none |
| 10 | Llama 3.18B Meta · mostly Q4_K_M Gen 1.5 t/s (1 run) | 1 | 1.5 1 run | no PP data | none |
| 11 | GLM-5.2744B · 40B active Zhipu AI · mostly fmt=4 grouped int4, gs=64 No generation data | 1 | no generation data | no PP data | none |
| 12 | Muse Glimmer30B Meta · mostly UD-Q2-K-XL No generation data · with speculation 20 t/s (1) | 1 | no generation data | no PP data | 20 1 run |
| 13 | Ornith1.59B mostly Q6_K No generation data | 1 | no generation data | no PP data | none |
| 14 | Qwen3.627B Alibaba · mostly Q4_K_M No generation data | 1 | no generation data | no PP data | none |
| 15 | Qwen3.6 Omnimerge-v427B Alibaba · mostly IQ3_M No generation data · with speculation 46 t/s (1) | 1 | no generation data | no PP data | 46 1 run |
| 16 | Qwen3.6 Tiel-Coder35B · 3B active Alibaba · mostly Q4_K_XL No generation data | 1 | no generation data | no PP data | none |
| 17 | Qwen3.8 ASCII-Condensed27B Alibaba · mostly UD-IQ4_XS No generation data | 1 | no generation data | no PP data | none |
| 18 | Qwen3.8 Huihui-Abliterated27B Alibaba · mostly IQ3_XXS No generation data | 1 | no generation data | no PP data | none |
Not in the table: 4 reports don't say which model size ran. 14 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.