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 q4_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 | 73.2 | 73.6 | 74.3 | 75.7 | 78.5 | 84.1 |
| Q6_K | 59.93 GiB | 61.2 | 61.6 | 62.3 | 63.7 | 66.5 | 72.1 |
| Q5_K_M | 50.71 GiB | 52.0 | 52.3 | 53.0 | 54.5 | 57.3 | 62.9 |
| Q4_K_M | 44.16 GiB | 45.4 | 45.8 | 46.5 | 47.9 | 50.7 | 56.3 |
| Q4_0 | 38.54 GiB | 39.8 | 40.2 | 40.9 | 42.3 | 45.1 | 50.7 |
| IQ4_XS | 36.98 GiB | 38.3 | 38.6 | 39.3 | 40.7 | 43.5 | 49.2 |
| Q3_K_L | 36.79 GiB | 38.1 | 38.4 | 39.1 | 40.5 | 43.3 | 49.0 |
| Q3_K_M | 35.11 GiB | 36.4 | 36.7 | 37.4 | 38.9 | 41.7 | 47.3 |
| IQ3_M | 33.07 GiB | 34.3 | 34.7 | 35.4 | 36.8 | 39.6 | 45.2 |
| Q3_K_S | 32.12 GiB | 33.4 | 33.8 | 34.5 | 35.9 | 38.7 | 44.3 |
| IQ3_XXS | 29.66 GiB | 30.9 | 31.3 | 32.0 | 33.4 | 36.2 | 41.8 |
| Q2_K_L | 28.90 GiB | 30.2 | 30.5 | 31.2 | 32.6 | 35.5 | 41.1 |
| Q2_K | 27.76 GiB | 29.0 | 29.4 | 30.1 | 31.5 | 34.3 | 39.9 |
| IQ2_M | 27.32 GiB | 28.6 | 29.0 | 29.7 | 31.1 | 33.9 | 39.5 |
| IQ2_XS | 25.20 GiB | 26.5 | 26.8 | 27.5 | 28.9 | 31.8 | 37.4 |
| IQ2_XXS | 23.74 GiB | 25.0 | 25.4 | 26.1 | 27.5 | 30.3 | 35.9 |
| IQ1_M | 22.11 GiB | 23.4 | 23.7 | 24.4 | 25.9 | 28.7 | 34.3 |
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.