Boards
16 measured pairs. Rank is ordinal tok/s inside this filter — not a composite score.
| # | Hardware | Model · Build | Runtime | Decode | Prefill | Peak | Get | Proof |
|---|---|---|---|---|---|---|---|---|
| 1 |  RTX 5090 32GB |  Qwen3.5 2BNewHot | unknown | 351 | — | — |  @fillagrew | |
| 2 | RTX 5090 32GB |  Gemma-4 26B-A4B-itNewHot | unknown | 173 | — | — | @fillagrew | |
| 3 |   M5 Max | Qwen3.8-Flash-NextNewHot |  MTPLX V2.11.3 | 126.5 | — | — | @Youssofal_ | |
| 4 |  2× DGX Spark | Qwen3.6-35B-A3BHot | unknown | 94.4 | 181.5 | — | @bonellisystems | |
| 5 |  M4 Pro | Qwen3.6-35B-A3B | rapid-mlx | 85.5 | — | 21 GB |  HF | @rapidmlx |
| 6 | M4 Pro | Qwen3.5-4BNew | rapid-mlx | 82.8 | — | 8 GB | @rapidmlx | |
| 7 |  RTX 3060 12GB | E Empero Qwen3.8-35B-A3BNewHot |  llama.cpp | 50 | 550 | — | @Oluwaphilemon1 | |
| 8 | M4 Pro | Qwen3.5-9BNew | rapid-mlx | 49.3 | — | — | @rapidmlx | |
| 9 | M4 Pro | Qwen3-8B | rapid-mlx | 48.3 | — | — | @rapidmlx | |
| 10 | 2× DGX Spark | Qwen3.8-Flash-NextNewHot | unknown | 45 | — | — | @Oluwaphilemon1 | |
| 11 |  RTX 4090 24GB | Qwen3.8-27BNewHot | llama.cpp | 40.7 | 2659.8 | 23.7 VRAM | HF | @Oluwaphilemon1 |
| 12 |  GTX 1660 SUPER 6GB | O OrnithNew | unknown | 40 | — | — | HF | @mine_craft_bui |
| 13 |  RTX 5060 Ti 16GB | Qwen3.8-27BNewHot | llama.cpp | 23 | — | — | HF | @fntAInhead |
| 14 | RTX 5060 Ti 16GB | Qwen3.8-27BNewHot | llama.cpp | 19 | — | — | HF | @fntAInhead |
| 15 | RTX 5060 Ti 16GB | N Nex-N2.5-miniNew | llama.cpp | 14 | — | — | @fntAInhead | |
| 16 | RTX 5060 Ti 16GB | Qwen3.8-27B TurboFCFusionNewHot | llama.cpp | 10 | — | — | @fntAInhead |