bestmodel.run / model — gemma-4-26B-A4B-it (agent view) category chat · google/gemma-4-26B-A4B-it · 47 runs in the model record answer: 63.3 tok/s (model median) Honesty ladder: measured > reported > extrapolated > formula > no data yet. Empty fields stay empty — never filled with an estimate. rig | quant | median result | basis | ttft | peak vram | max context Arc Pro B70 32GB | 8-bit | 43.3 tok/s | reported | ttft 393ms | vram - GB | ctx 2,048 B200 192GB | 16-bit | 245.3 tok/s | reported | ttft 20ms | vram 161.2 GB | ctx 32,768 AMD Ryzen 7 8745H w/ Radeon 780M | 4-bit | 30.6 tok/s | reported | ttft 1,534ms | vram 4 GB | ctx 256,000 M1 Max 64GB | 8-bit | 49 tok/s | reported | ttft 462ms | vram - GB | ctx 2,048 M3 Ultra 512GB | 4-bit | 76.9 tok/s | reported | ttft 279ms | vram - GB | ctx 2,048 Radeon AI Pro R9700 32GB ×2 | 4-bit | 87.3 tok/s | reported | ttft 375ms | vram - GB | ctx 794 Radeon AI Pro R9700 32GB ×2 | 8-bit | 99.3 tok/s | measured | ttft -ms | vram - GB | ctx 262,144 Radeon AI Pro R9700 32GB ×3 | 8-bit | 75.3 tok/s | measured | ttft 322ms | vram 35.6 GB | ctx 262,144 RTX 3090 24GB ×2 | 8-bit | 154.1 tok/s | reported | ttft 42ms | vram 45.3 GB | ctx 262,144 RTX 4070 12GB | 4-bit | 20.4 tok/s | measured | ttft 1,661ms | vram - GB | ctx 2,048 RTX 5090 32GB | 8-bit | 282 tok/s | measured | ttft 30ms | vram - GB | ctx 2,048 RTX PRO 6000 Blackwell 96GB | 8-bit | 276.7 tok/s | reported | ttft 28ms | vram - GB | ctx 8,192 RX 570 4GB ×2 | 4-bit | 12.5 tok/s | measured | ttft 760ms | vram - GB | ctx 2,048 RX 7900 XTX 24GB ×2 | 8-bit | 99.6 tok/s | reported | ttft 282ms | vram - GB | ctx 8,192 Ryzen AI Max 395 128GB | 4-bit | 60.7 tok/s | measured | ttft -ms | vram 21.5 GB | ctx 32,768 Ryzen AI Max 395 128GB | 6-bit | 55.5 tok/s | reported | ttft -ms | vram 23.8 GB | ctx 16,384 Ryzen AI Max 395 128GB | 8-bit | 46.1 tok/s | reported | ttft 746ms | vram 27.8 GB | ctx 16,384 Tesla P100-PCIE-16GB 16GB ×2 | 5-bit | 54.8 tok/s | reported | ttft -ms | vram - GB | ctx 2,048