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.
Strix Halo and DGX Spark· NVIDIA only
128 GB of unified memory at about 260 GB/s: big models fit, and speed is set by that bandwidth. Memory speed 273 GB/s across the 2 devices with reports, so a typical figure here spans all of them. 52 single-device reports.
| # | Model | Reports | Gen t/s | PP t/s | With speculation |
|---|---|---|---|---|---|
| 1 | Qwen3.8 Flash-Next125B · 6B active Alibaba · mostly NVFP4 Gen 35 t/s (3 runs) · PP 899 t/s (1 run) · with speculation 30 t/s (3) | 11 | 35 3 runs, 2 devices | 899 1 run | 30 3 runs |
| 2 | Qwen3.827B Alibaba · mostly NVFP4 No generation data · PP 900 t/s (1 run) · with speculation 40 t/s (6) | 10 | no generation data | 900 1 run | 40 6 runs |
| 3 | DeepSeek V4 Flash284B · 13B active DeepSeek · mostly 2-bit No generation data · with speculation 17 t/s (2) | 9 | no generation data | no PP data | 17 2 runs |
| 4 | Qwen3.635B · 3B active Alibaba · mostly NVFP4 Gen 78 t/s (1 run) · PP 3557 t/s (1 run) · with speculation 139 t/s (1) | 3 | 78 1 run | 3557 1 run | 139 1 run |
| 5 | Ling-3.0 Flash124B · 5.1B active Ant Group · mostly Q5_K_M Gen 39 t/s (3 runs) | 3 | 39 3 runs | no PP data | none |
| 6 | Ornith1.535B · 3B active mostly NVFP4 Gen 69 t/s (1 run) · with speculation 79 t/s (1) | 2 | 69 1 run | no PP data | 79 1 run |
| 7 | Gemma 426B · 4B active Google DeepMind · mostly NVFP4 Gen 59 t/s (2 runs) · PP 2657 t/s (1 run) | 2 | 59 2 runs | 2657 1 run | none |
| 8 | Nemotron 3 Super120B · 12B active mostly Q4_K Gen 17 t/s (2 runs) · PP 769 t/s (1 run) | 2 | 17 2 runs | 769 1 run | none |
| 9 | DeepSeek V4.1 Flash552B · 16B active mostly FP4 No generation data | 2 | no generation data | no PP data | none |
| 10 | Qwen3.627B Alibaba · mostly NVFP4 No generation data | 2 | no generation data | no PP data | none |
| 11 | gpt-oss117B · 5.1B active OpenAI · mostly MXFP4 Gen 53 t/s (1 run) · PP 1689 t/s (1 run) | 1 | 53 1 run | 1689 1 run | none |
| 12 | DeepSeek V4 Flash162B · 13B active DeepSeek · mostly NVFP4 Gen 14 t/s (1 run) · PP 539 t/s (1 run) | 1 | 14 1 run | 539 1 run | none |
| 13 | Gemma 431B Google DeepMind Gen 6.0 t/s (1 run) | 1 | 6.0 1 run | no PP data | none |
| 14 | Gemma 4 E4B8B Google DeepMind · mostly Q4_K_M No generation data | 1 | no generation data | no PP data | none |
| 15 | Mimo 2.6 Pro1020B · 42B active Xiaomi · mostly FP8 No generation data | 1 | no generation data | no PP data | none |
| 16 | Nemotron 3.5 Lightning30B · 3B active NVIDIA · mostly Q4_K_M No generation data | 1 | no generation data | no PP data | none |
Not in the table: 4 reports don't say which model size ran. 42 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.