Does Llama-4-Scout-17B-16E-Instruct-abliterated-v2 fit in 12GB of VRAM?
Not at these settings. No indexed quantization of Llama-4-Scout-17B-16E-Instruct-abliterated-v2 fits 12GB card at any context we compute, with q8_0 KV. The smallest shipped quantization is 20.66 GiB in weights alone, against 11.16 GiB usable. CPU offload can still run it, slowly.
Every quantization at every context
| Quant | Weights● | 4K◐ | 8K◐ | 16K◐ | 32K◐ | 64K◐ | 128K◐ |
|---|---|---|---|---|---|---|---|
| I1-IQ3_M | 44.25 GiB | 45.5 | 45.9 | 46.7 | 48.3 | 51.4 | 57.8 |
| I1-IQ3_S | 43.56 GiB | 44.8 | 45.2 | 46.0 | 47.6 | 50.8 | 57.1 |
| I1-Q3_K_S | 43.53 GiB | 44.8 | 45.2 | 46.0 | 47.5 | 50.7 | 57.1 |
| Q3_K_S | 43.53 GiB | 44.8 | 45.2 | 46.0 | 47.5 | 50.7 | 57.1 |
| I1-IQ3_XS | 41.25 GiB | 42.5 | 42.9 | 43.7 | 45.3 | 48.5 | 54.8 |
| I1-IQ3_XXS | 38.80 GiB | 40.0 | 40.4 | 41.2 | 42.8 | 46.0 | 52.4 |
| I1-Q2_K | 36.85 GiB | 38.1 | 38.5 | 39.3 | 40.9 | 44.0 | 50.4 |
| Q2_K | 36.85 GiB | 38.1 | 38.5 | 39.3 | 40.9 | 44.0 | 50.4 |
| I1-Q2_K_S | 34.42 GiB | 35.6 | 36.0 | 36.8 | 38.4 | 41.6 | 48.0 |
| I1-IQ2_M | 33.04 GiB | 34.3 | 34.7 | 35.5 | 37.1 | 40.2 | 46.6 |
| I1-IQ2_S | 30.07 GiB | 31.3 | 31.7 | 32.5 | 34.1 | 37.3 | 43.6 |
| I1-IQ2_XS | 29.60 GiB | 30.8 | 31.2 | 32.0 | 33.6 | 36.8 | 43.2 |
| I1-IQ2_XXS | 26.60 GiB | 27.8 | 28.2 | 29.0 | 30.6 | 33.8 | 40.2 |
| I1-IQ1_M | 22.88 GiB | 24.1 | 24.5 | 25.3 | 26.9 | 30.1 | 36.5 |
| I1-IQ1_S | 20.66 GiB | 21.9 | 22.3 | 23.1 | 24.7 | 27.9 | 34.2 |
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 12GB 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 8,192-token window rather than the full context.