Does Step-3.7-Flash fit in 32GB of VRAM?
Not at these settings. No indexed quantization of Step-3.7-Flash fits 32GB card at any context we compute, with q4_0 KV. The smallest shipped quantization is 40.03 GiB in weights alone, against 29.76 GiB usable. CPU offload can still run it, slowly.
Every quantization at every context
| Quant | Weights● | 4K◐ | 8K◐ | 16K◐ | 32K◐ | 64K◐ | 128K◐ |
|---|---|---|---|---|---|---|---|
| BF16 | 366.96 GiB | 368.5 | 368.9 | 369.8 | 371.5 | 374.8 | 381.6 |
| Q8_0 | 197.44 GiB | 199.0 | 199.4 | 200.2 | 201.9 | 205.3 | 212.1 |
| Q6_K_L | 160.20 GiB | 161.7 | 162.2 | 163.0 | 164.7 | 168.1 | 174.8 |
| Q6_K | 159.96 GiB | 161.5 | 161.9 | 162.8 | 164.5 | 167.8 | 174.6 |
| UD-Q6_K | 152.12 GiB | 153.7 | 154.1 | 154.9 | 156.6 | 160.0 | 166.7 |
| UD-Q5_K_M | 136.43 GiB | 138.0 | 138.4 | 139.2 | 140.9 | 144.3 | 151.1 |
| Q5_K_L | 132.58 GiB | 134.1 | 134.5 | 135.4 | 137.1 | 140.5 | 147.2 |
| Q5_K_M | 132.28 GiB | 133.8 | 134.2 | 135.1 | 136.8 | 140.1 | 146.9 |
| UD-Q5_K_S | 128.59 GiB | 130.1 | 130.6 | 131.4 | 133.1 | 136.5 | 143.2 |
| Q5_K_S | 128.01 GiB | 129.6 | 130.0 | 130.8 | 132.5 | 135.9 | 142.6 |
| Q4_1 | 116.67 GiB | 118.2 | 118.6 | 119.5 | 121.2 | 124.5 | 131.3 |
| UD-Q4_K_M | 113.71 GiB | 115.2 | 115.7 | 116.5 | 118.2 | 121.6 | 128.3 |
| Q4_K_L | 113.63 GiB | 115.2 | 115.6 | 116.4 | 118.1 | 121.5 | 128.3 |
| Q4_K_M | 113.27 GiB | 114.8 | 115.2 | 116.1 | 117.8 | 121.1 | 127.9 |
| Q4_K_S | 109.05 GiB | 110.6 | 111.0 | 111.9 | 113.5 | 116.9 | 123.7 |
| UD-Q4_K_S | 106.32 GiB | 107.9 | 108.3 | 109.1 | 110.8 | 114.2 | 120.9 |
| Q4_0 | 105.79 GiB | 107.3 | 107.8 | 108.6 | 110.3 | 113.7 | 120.4 |
| IQ4_NL | 105.54 GiB | 107.1 | 107.5 | 108.3 | 110.0 | 113.4 | 120.2 |
| IQ4_XS | 99.94 GiB | 101.5 | 101.9 | 102.7 | 104.4 | 107.8 | 114.6 |
| UD-IQ4_NL | 90.63 GiB | 92.2 | 92.6 | 93.4 | 95.1 | 98.5 | 105.3 |
| IQ3_M | 89.38 GiB | 90.9 | 91.3 | 92.2 | 93.9 | 97.3 | 104.0 |
| Q3_K_L | 88.94 GiB | 90.5 | 90.9 | 91.7 | 93.4 | 96.8 | 103.6 |
| UD-IQ4_XS | 88.79 GiB | 90.3 | 90.8 | 91.6 | 93.3 | 96.7 | 103.4 |
| Q3_K_M | 87.36 GiB | 88.9 | 89.3 | 90.2 | 91.9 | 95.2 | 102.0 |
| IQ3_XS | 85.44 GiB | 87.0 | 87.4 | 88.2 | 89.9 | 93.3 | 100.1 |
| UD-Q3_K_M | 83.13 GiB | 84.7 | 85.1 | 85.9 | 87.6 | 91.0 | 97.8 |
| Q3_K_S | 81.55 GiB | 83.1 | 83.5 | 84.4 | 86.0 | 89.4 | 96.2 |
| IQ3_XXS | 78.38 GiB | 79.9 | 80.3 | 81.2 | 82.9 | 86.3 | 93.0 |
| UD-IQ3_S | 74.54 GiB | 76.1 | 76.5 | 77.3 | 79.0 | 82.4 | 89.2 |
| UD-IQ3_XXS | 68.54 GiB | 70.1 | 70.5 | 71.3 | 73.0 | 76.4 | 83.2 |
| Q2_K_L | 66.82 GiB | 68.4 | 68.8 | 69.6 | 71.3 | 74.7 | 81.4 |
| Q2_K | 66.34 GiB | 67.9 | 68.3 | 69.1 | 70.8 | 74.2 | 81.0 |
| IQ2_M | 63.68 GiB | 65.2 | 65.6 | 66.5 | 68.2 | 71.6 | 78.3 |
| IQ2_S | 57.93 GiB | 59.5 | 59.9 | 60.7 | 62.4 | 65.8 | 72.5 |
| UD-IQ2_M | 57.58 GiB | 59.1 | 59.5 | 60.4 | 62.1 | 65.4 | 72.2 |
| UD-IQ2_XXS | 57.47 GiB | 59.0 | 59.4 | 60.3 | 62.0 | 65.3 | 72.1 |
| IQ2_XS | 56.76 GiB | 58.3 | 58.7 | 59.6 | 61.2 | 64.6 | 71.4 |
| UD-IQ1_M | 52.85 GiB | 54.4 | 54.8 | 55.7 | 57.3 | 60.7 | 67.5 |
| IQ2_XXS | 51.22 GiB | 52.8 | 53.2 | 54.0 | 55.7 | 59.1 | 65.8 |
| IQ1_M | 44.41 GiB | 45.9 | 46.4 | 47.2 | 48.9 | 52.3 | 59.0 |
| IQ1_S | 40.03 GiB | 41.6 | 42.0 | 42.8 | 44.5 | 47.9 | 54.7 |
Figures are GiB of total memory: weights plus KV cache plus compute buffer and backend overhead. Weights and KV are near-exact; the overhead term is modeled. Hover any cell for the breakdown.
Why there are no speeds on this page
A capacity is not a card. Whether a model fits depends only on memory, so every figure above holds for any 32GB accelerator. How fast it runs depends on memory bandwidth, which varies several-fold between cards of the same capacity — so putting a tokens-per-second number here would be inventing one. Pick a specific card from hardware and the speed column appears.
Why other calculators disagree
A parameters × bits ÷ 8 estimate ignores two things that dominate at long context. First, the weights themselves are not the nominal rate — quantizations are mixtures, so the real file is consistently larger than the label implies. Second, most of this model's layers cache only a 512-token window rather than the full context.