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 8-16 GB· Apple only
Base Macs. Small models only, since macOS keeps a share of memory for itself. Memory speed 68 to 205 GB/s across the 7 devices with reports, so a typical figure here spans all of them. 32 single-device reports.
| # | Model | Reports | Gen t/s | PP t/s | With speculation |
|---|---|---|---|---|---|
| 1 | Qwen3.635B · 3B active Alibaba · mostly 4-bit Gen 38 t/s (1 run) · PP 274 t/s (1 run) | 3 | 38 1 run | 274 1 run | none |
| 2 | Gemma 426B · 4B active Google DeepMind · mostly IQ4_XS Gen 31 t/s (1 run) | 2 | 31 1 run | no PP data | none |
| 3 | Qwen3.54B Alibaba · mostly Q4_K_M Gen 28 t/s (1 run) · PP 352 t/s (1 run) | 2 | 28 1 run | 352 1 run | none |
| 4 | Qwen3.535B · 3B active Alibaba Gen 17 t/s (1 run) | 2 | 17 1 run | no PP data | none |
| 5 | Qwen38B Alibaba · mostly 4-bit Gen 17 t/s (2 runs) · PP 219 t/s (1 run) | 2 | 17 2 runs | 219 1 run | none |
| 6 | Qwen314B Alibaba · mostly 4-bit Gen 15 t/s (2 runs) · PP 116 t/s (1 run) | 2 | 15 2 runs, 2 devices | 116 1 run | none |
| 7 | Qwen3.827B Alibaba · mostly Q3_xxs Gen 9.0 t/s (1 run) | 2 | 9.0 1 run | no PP data | none |
| 8 | Llama 3.23B Meta · mostly Q4_K_M No generation data | 2 | no generation data | no PP data | none |
| 9 | Qwen3.59B Alibaba · mostly Q4_K_M No generation data | 2 | no generation data | no PP data | none |
| 10 | Qwen2.50.5B Alibaba · mostly Q4 Gen 96 t/s (1 run) | 1 | 96 1 run | no PP data | none |
| 11 | Qwen3.50.8B Alibaba · mostly Q8_0 Gen 90 t/s (1 run) · PP 2122 t/s (1 run) | 1 | 90 1 run | 2122 1 run | none |
| 12 | Mellum2.1 Thinking12B · 2.5B active mostly NVFP4 Gen 64 t/s (1 run) | 1 | 64 1 run | no PP data | none |
| 13 | Qwen2.53B Alibaba · mostly Q4_K_M Gen 47 t/s (1 run) · PP 557 t/s (1 run) | 1 | 47 1 run | 557 1 run | none |
| 14 | LFM2.52.6B Liquid AI · mostly QAD-Q4_0 Gen 22 t/s (1 run) · PP 200 t/s (1 run) | 1 | 22 1 run | 200 1 run | none |
| 15 | Gemma 4 E4B8B Google DeepMind · mostly 4bit Gen 8.2 t/s (1 run) | 1 | 8.2 1 run | no PP data | none |
| 16 | Bonsai 227B PrismML · mostly ternary Gen 7.6 t/s (1 run) · PP 42 t/s (1 run) | 1 | 7.6 1 run | 42 1 run | none |
| 17 | Gemma 4 E2B5.1B Google DeepMind · mostly Q4 No generation data | 1 | no generation data | no PP data | none |
| 18 | Gemma 412B Google DeepMind · mostly Q4_K_M No generation data | 1 | no generation data | no PP data | none |
| 19 | Llama 3.18B Meta · mostly Q4_K_M No generation data | 1 | no generation data | no PP data | none |
| 20 | Mellum2.112B · 2.5B active mostly Q8 No generation data | 1 | no generation data | no PP data | none |
| 21 | Qwen2.51.5B Alibaba · mostly Q2_K No generation data | 1 | no generation data | no PP data | none |
| 22 | Qwen3.8 Flash-Next125B · 6B active Alibaba · mostly 4-bit No generation data | 1 | no generation data | no PP data | none |
Not in the table: 8 reports don't say which model size ran. 3 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.