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.
A Mac counts when its report names the product, or when only that product was sold with its chip and memory.
Apple Silicon 18-36 GB· Mac mini only
Mid-size Macs. Up to about 30B-class models at 4 bits. Memory speed 102 to 273 GB/s across the 7 devices with reports, so a typical figure here spans all of them. 43 single-device reports.
| # | Model | Reports | Gen t/s | PP t/s | With speculation |
|---|---|---|---|---|---|
| 1 | Qwen3.827B Alibaba · mostly Q4_K_M Gen 6.5 t/s (1 run) · PP 61 t/s (1 run) · with speculation 27 t/s (1) | 5 | 6.5 1 run | 61 1 run | 27 1 run |
| 2 | Gemma 412B Google DeepMind · mostly Q4_K_M Gen 13 t/s (1 run) · PP 139 t/s (1 run) | 4 | 13 1 run | 139 1 run | none |
| 3 | Qwen3.8 Flash-Next125B · 6B active Alibaba · mostly Q2_0 No generation data | 4 | no generation data | no PP data | none |
| 4 | Gemma 426B · 4B active Google DeepMind · mostly 4-bit Gen 52 t/s (1 run) | 3 | 52 1 run | no PP data | none |
| 5 | Qwen3.59B Alibaba · mostly Q4_K_M Gen 17 t/s (1 run) | 3 | 17 1 run | no PP data | none |
| 6 | Gemma 4 E4B8B Google DeepMind · mostly Q4_K_M No generation data | 3 | no generation data | no PP data | none |
| 7 | Qwen3.635B · 3B active Alibaba · mostly 4-bit group-64 Gen 64 t/s (1 run) | 2 | 64 1 run | no PP data | none |
| 8 | Gemma 431B Google DeepMind · mostly Q4_K_M No generation data | 2 | no generation data | no PP data | none |
| 9 | Nemotron 3.5 Lightning30B · 3B active NVIDIA · mostly Q4_K_M No generation data | 2 | no generation data | no PP data | none |
| 10 | Llama 3.23B Meta · mostly Q4_K_M Gen 46 t/s (1 run) · PP 530 t/s (1 run) | 1 | 46 1 run | 530 1 run | none |
| 11 | Qwen34B Alibaba · mostly Q4_K_M Gen 36 t/s (1 run) · PP 417 t/s (1 run) | 1 | 36 1 run | 417 1 run | none |
| 12 | Qwen 2.5 Coder7B mostly Q4_K_M Gen 23 t/s (1 run) · PP 240 t/s (1 run) | 1 | 23 1 run | 240 1 run | none |
| 13 | Qwen38B Alibaba · mostly Q4_K_M Gen 20 t/s (1 run) · PP 224 t/s (1 run) | 1 | 20 1 run | 224 1 run | none |
| 14 | Qwen3.54B Alibaba · mostly affine Q4, group size 64 Gen 16 t/s (1 run) · PP 439 t/s (1 run) | 1 | 16 1 run | 439 1 run | none |
| 15 | Qwen314B Alibaba · mostly 4-bit Gen 12 t/s (1 run) | 1 | 12 1 run | no PP data | none |
| 16 | Qwen3.8 27B dense27B Alibaba · mostly 4bit Gen 9.8 t/s (1 run) | 1 | 9.8 1 run | no PP data | none |
| 17 | Gemma 327B Google · mostly 4-bit Gen 8.9 t/s (1 run) · PP 11 t/s (1 run) | 1 | 8.9 1 run | 11 1 run | none |
| 18 | Gemma 4 E2B5.1B Google DeepMind · mostly Q4_K_M No generation data | 1 | no generation data | no PP data | none |
| 19 | Ornith1.535B · 3B active mostly Q4_K_M No generation data | 1 | no generation data | no PP data | none |
| 20 | Ornith1.5 Ornith1.5-35B-A3B Splash35B · 3B active No generation data | 1 | no generation data | no PP data | none |
| 21 | Qwen30.6B Alibaba · mostly Q4 No generation data | 1 | no generation data | no PP data | none |
| 22 | Qwen3.8 Swift-Qwen3.8-27B Splash27B Alibaba No generation data | 1 | no generation data | no PP data | none |
| 23 | Qwen3.8 Qwen3.8-27B Splash27B Alibaba No generation data | 1 | no generation data | no PP data | none |
| 24 | gpt-oss21B · 3.6B active OpenAI No generation data | 1 | no generation data | no PP data | none |
Not in the table: 11 reports don't say which model size ran.
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.