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 128 GB+· Mac Studio only
Max and Ultra Macs with room for large mixture-of-experts models. Memory speed 546 to 1,229 GB/s across the 9 devices with reports, so a typical figure here spans all of them. 121 single-device reports.
| # | Model | Reports | Gen t/s | PP t/s | With speculation |
|---|---|---|---|---|---|
| 1 | Qwen3.827B Alibaba · mostly 4bit Gen 23 t/s (14 runs) · PP 250 t/s (13 runs) · with speculation 53 t/s (3) | 20 | 23 14 runs, 6 devices | 250 13 runs | 53 3 runs |
| 2 | Qwen3.8 Flash-Next125B · 6B active Alibaba · mostly oQ4e Gen 27 t/s (8 runs) · PP 1069 t/s (3 runs) · with speculation 104 t/s (8) | 19 | 27 8 runs, 4 devices | 1069 3 runs | 104 8 runs |
| 3 | GLM-5.3 Flash320B · 18B active Zhipu AI · mostly oQ4e Gen 57 t/s (2 runs) · PP 550 t/s (1 run) · with speculation 70 t/s (4) | 10 | 57 2 runs, 2 devices | 550 1 run | 70 4 runs |
| 4 | DeepSeek V4 Flash284B · 13B active DeepSeek · mostly MXFP4 Gen 22 t/s (3 runs) · PP 451 t/s (4 runs) · with speculation 15 t/s (1) | 7 | 22 3 runs, 2 devices | 451 4 runs | 15 1 run |
| 5 | Gemma 426B · 4B active Google DeepMind · mostly 4bit MLX Gen 88 t/s (5 runs) · PP 2595 t/s (5 runs) | 6 | 88 5 runs, 2 devices | 2595 5 runs | none |
| 6 | Qwen3.59B Alibaba · mostly Q4_K_M Gen 68 t/s (3 runs) · PP 856 t/s (2 runs) | 5 | 68 3 runs, 2 devices | 856 2 runs | none |
| 7 | DeepSeek V4.1 Flash552B · 16B active mostly oQ4e Gen 45 t/s (2 runs) · PP 655 t/s (2 runs) · with speculation 31 t/s (1) | 5 | 45 2 runs, 2 devices | 655 2 runs | 31 1 run |
| 8 | Qwen3.635B · 3B active Alibaba · mostly Q4 Gen 81 t/s (2 runs) · PP 1874 t/s (2 runs) | 4 | 81 2 runs, 2 devices | 1874 2 runs | none |
| 9 | gpt-oss117B · 5.1B active OpenAI · mostly MXFP4 Gen 75 t/s (2 runs) · PP 601 t/s (2 runs) | 3 | 75 2 runs, 2 devices | 601 2 runs | none |
| 10 | Qwen3.5122B · 10B active Alibaba · mostly Q4 Gen 50 t/s (3 runs) · PP 1748 t/s (2 runs) | 3 | 50 3 runs, 3 devices | 1748 2 runs | none |
| 11 | Gemma 431B Google DeepMind · mostly BF16 Gen 12 t/s (2 runs) · PP 766 t/s (2 runs) | 3 | 12 2 runs, 2 devices | 766 2 runs | none |
| 12 | Qwen314B Alibaba · mostly Q4 No generation data | 3 | no generation data | no PP data | none |
| 13 | SmolLM33B mostly Q4_K_M Gen 144 t/s (2 runs) · PP 2190 t/s (2 runs) | 2 | 144 2 runs | 2190 2 runs | none |
| 14 | Qwen3.54B Alibaba · mostly Q4_K_M Gen 90 t/s (2 runs) · PP 1522 t/s (2 runs) | 2 | 90 2 runs | 1522 2 runs | none |
| 15 | Qwen3-Coder30B · 3B active mostly UD-Q4_K_XL Gen 70 t/s (2 runs) · PP 1089 t/s (2 runs) | 2 | 70 2 runs | 1089 2 runs | none |
| 16 | Gemma 4 E4B8B Google DeepMind · mostly Q4_K_M Gen 70 t/s (1 run) · PP 1173 t/s (1 run) | 2 | 70 1 run | 1173 1 run | none |
| 17 | Qwen2.572B Alibaba · mostly 4-bit Gen 11 t/s (2 runs) | 2 | 11 2 runs, 2 devices | no PP data | none |
| 18 | Qwen3.5397B · 17B active Alibaba · mostly 4bit MLX No generation data | 2 | no generation data | no PP data | none |
| 19 | Falcon-H11.5B mostly Q4_K_M Gen 148 t/s (1 run) · PP 2535 t/s (1 run) | 1 | 148 1 run | 2535 1 run | none |
| 20 | Qwen3 30B-A3B, standard 8 experts30B · 3.3B active Alibaba · mostly MLX 4-bit affine, group 64 Gen 144 t/s (1 run) | 1 | 144 1 run | no PP data | none |
| 21 | Granite-4.0-H1B mostly Q4_K_M Gen 142 t/s (1 run) · PP 2656 t/s (1 run) | 1 | 142 1 run | 2656 1 run | none |
| 22 | gpt-oss21B · 3.6B active OpenAI · mostly MXFP4 Gen 123 t/s (1 run) · PP 1577 t/s (1 run) | 1 | 123 1 run | 1577 1 run | none |
| 23 | Qwen34B Alibaba · mostly 4-bit Gen 117 t/s (1 run) · PP 942 t/s (1 run) | 1 | 117 1 run | 942 1 run | none |
| 24 | Falcon-H13B mostly Q4_K_M Gen 86 t/s (1 run) · PP 1298 t/s (1 run) | 1 | 86 1 run | 1298 1 run | none |
| 25 | Qwen38B Alibaba · mostly 4-bit Gen 81 t/s (1 run) · PP 518 t/s (1 run) | 1 | 81 1 run | 518 1 run | none |
| 26 | Qwen3.535B · 3B active Alibaba · mostly MXFP4_MOE Gen 81 t/s (1 run) · PP 2779 t/s (1 run) | 1 | 81 1 run | 2779 1 run | none |
| 27 | Ling-3.0 Flash124B · 5.1B active Ant Group · mostly oQ5 affine mixed precision, group 64 Gen 73 t/s (1 run) | 1 | 73 1 run | no PP data | none |
| 28 | Gemma 412B Google DeepMind · mostly Q4_0 Gen 55 t/s (1 run) · PP 589 t/s (1 run) | 1 | 55 1 run | 589 1 run | none |
| 29 | Falcon-H17B mostly Q4_K_M Gen 53 t/s (1 run) · PP 683 t/s (1 run) | 1 | 53 1 run | 683 1 run | none |
| 30 | Bonsai 2 Ternary27B PrismML · mostly PQ2_0 Gen 34 t/s (1 run) | 1 | 34 1 run | no PP data | none |
| 31 | MiniMax M2.7230B · 10B active mostly UD-Q4_K_M Gen 33 t/s (1 run) · PP 421 t/s (1 run) | 1 | 33 1 run | 421 1 run | none |
| 32 | Qwen332B Alibaba · mostly Q4 Gen 31 t/s (1 run) | 1 | 31 1 run | no PP data | none |
| 33 | Minimax M2.1230B · 10B active mostly 4bit Gen 25 t/s (1 run) | 1 | 25 1 run | no PP data | none |
| 34 | Muse Spark Open816B · 42B active Meta Gen 4.0 t/s (1 run) | 1 | 4.0 1 run | no PP data | none |
| 35 | Llama 3.1405B Meta · mostly Q4 Gen 3.0 t/s (1 run) | 1 | 3.0 1 run | no PP data | none |
| 36 | DeepSeek V4 Flash Vision-Exp285B · 13B active DeepSeek · mostly IQ2_XXS/Q2_K No generation data · with speculation 32 t/s (1) | 1 | no generation data | no PP data | 32 1 run |
| 37 | Llama 3.370B Meta · mostly Q8 No generation data | 1 | no generation data | no PP data | none |
| 38 | Mimo 2.6 Flash309B · 15B active Xiaomi No generation data | 1 | no generation data | no PP data | none |
| 39 | Qwen330B · 3B active Alibaba No generation data | 1 | no generation data | no PP data | none |
Not in the table: 20 reports don't say which model size ran. 7 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.