Does Qwen2.5-Math-72B-Instruct fit in 8GB of VRAM?
Not at these settings. No indexed quantization of Qwen2.5-Math-72B-Instruct fits 8GB card at any context we compute, with q8_0 KV. The smallest shipped quantization is 22.11 GiB in weights alone, against 7.44 GiB usable. CPU offload can still run it, slowly.
Every quantization at every context
| Quant | Weights● | 4K◐ | 8K◐ | 16K◐ | 32K◐ | 64K◐ | 128K◐ |
|---|---|---|---|---|---|---|---|
| Q8_0 | 71.96 GiB | 73.6 | 74.2 | 75.5 | 78.2 | 83.5 | 94.1 |
| Q6_K | 59.93 GiB | 61.5 | 62.2 | 63.5 | 66.2 | 71.5 | 82.1 |
| Q5_K_M | 50.71 GiB | 52.3 | 53.0 | 54.3 | 57.0 | 62.3 | 72.9 |
| Q4_K_M | 44.16 GiB | 45.8 | 46.4 | 47.7 | 50.4 | 55.7 | 66.3 |
| Q4_0 | 38.54 GiB | 40.1 | 40.8 | 42.1 | 44.8 | 50.1 | 60.7 |
| IQ4_XS | 36.98 GiB | 38.6 | 39.2 | 40.6 | 43.2 | 48.5 | 59.2 |
| Q3_K_L | 36.79 GiB | 38.4 | 39.1 | 40.4 | 43.0 | 48.3 | 59.0 |
| Q3_K_M | 35.11 GiB | 36.7 | 37.4 | 38.7 | 41.4 | 46.7 | 57.3 |
| IQ3_M | 33.07 GiB | 34.7 | 35.3 | 36.7 | 39.3 | 44.6 | 55.2 |
| Q3_K_S | 32.12 GiB | 33.7 | 34.4 | 35.7 | 38.4 | 43.7 | 54.3 |
| IQ3_XXS | 29.66 GiB | 31.3 | 31.9 | 33.2 | 35.9 | 41.2 | 51.8 |
| Q2_K_L | 28.90 GiB | 30.5 | 31.2 | 32.5 | 35.1 | 40.5 | 51.1 |
| Q2_K | 27.76 GiB | 29.4 | 30.0 | 31.4 | 34.0 | 39.3 | 49.9 |
| IQ2_M | 27.32 GiB | 28.9 | 29.6 | 30.9 | 33.6 | 38.9 | 49.5 |
| IQ2_XS | 25.20 GiB | 26.8 | 27.5 | 28.8 | 31.4 | 36.8 | 47.4 |
| IQ2_XXS | 23.74 GiB | 25.3 | 26.0 | 27.3 | 30.0 | 35.3 | 45.9 |
| IQ1_M | 22.11 GiB | 23.7 | 24.4 | 25.7 | 28.4 | 33.7 | 44.3 |
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 8GB 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 4,096-token window rather than the full context.