Can I run Moonlight-16B-A3B-Instruct on a GeForce RTX 3050?
Not at these settings. No indexed quantization of Moonlight-16B-A3B-Instruct fits GeForce RTX 3050 at any context we compute, with q8_0 KV. The smallest shipped quantization is 4.77 GiB in weights alone, against 5.58 GiB usable. CPU offload can still run it, slowly.
Every quantization at every context
| Quant | Weights● | 4K◐ | 8K◐ | 16K◐ | 32K◐ | 64K◐ | 128K◐ |
|---|---|---|---|---|---|---|---|
| F16 | 29.74 GiB | 30.6 | 30.7 | 30.8 | 31.1 | 31.6 | 32.6 |
| Q8_0 | 15.81 GiB | 16.7 | 16.7 | 16.9 | 17.1 | 17.6 | 18.6 |
| Q6_K | 13.30 GiB | 14.2 | 14.2 | 14.4 | 14.6 | 15.1 | 16.1 |
| Q5_K_M | 11.21 GiB | 12.1 | 12.2 | 12.3 | 12.5 | 13.0 | 14.0 |
| Q5_K_S | 10.56 GiB | 11.4 | 11.5 | 11.6 | 11.9 | 12.4 | 13.4 |
| Q5_0 | 10.30 GiB | 11.2 | 11.2 | 11.4 | 11.6 | 12.1 | 13.1 |
| Q4_K_M | 9.82 GiB | 10.7 | 10.8 | 10.9 | 11.1 | 11.6 | 12.6 |
| Q4_K_S | 9.05 GiB | 9.9 | 10.0 | 10.1 | 10.4 | 10.9 | 11.9 |
| Q4_0 | 8.48 GiB | 9.4 | 9.4 | 9.5 | 9.8 | 10.3 | 11.3 |
| IQ4_NL | 8.46 GiB | 9.3 | 9.4 | 9.5 | 9.8 | 10.3 | 11.3 |
| IQ4_XS | 8.15 GiB | 9.0 | 9.1 | 9.2 | 9.5 | 10.0 | 11.0 |
| Q3_K_L | 8.03 GiB | 8.9 | 9.0 | 9.1 | 9.3 | 9.9 | 10.9 |
| Q3_K_M | 7.72 GiB | 8.6 | 8.7 | 8.8 | 9.0 | 9.5 | 10.6 |
| IQ3_M | 7.19 GiB | 8.1 | 8.1 | 8.3 | 8.5 | 9.0 | 10.0 |
| IQ3_S | 7.12 GiB | 8.0 | 8.1 | 8.2 | 8.4 | 8.9 | 10.0 |
| Q3_K_S | 7.12 GiB | 8.0 | 8.1 | 8.2 | 8.4 | 8.9 | 10.0 |
| IQ3_XS | 6.78 GiB | 7.7 | 7.7 | 7.9 | 8.1 | 8.6 | 9.6 |
| IQ3_XXS | 6.62 GiB | 7.5 | 7.6 | 7.7 | 7.9 | 8.4 | 9.5 |
| Q2_K_S | 6.15 GiB | 7.0 | 7.1 | 7.2 | 7.5 | 8.0 | 9.0 |
| Q2_K | 6.13 GiB | 7.0 | 7.1 | 7.2 | 7.4 | 8.0 | 9.0 |
| IQ2_M | 6.03 GiB | 6.9 | 7.0 | 7.1 | 7.3 | 7.9 | 8.9 |
| IQ2_S | 5.73 GiB | 6.6 | 6.7 | 6.8 | 7.0 | 7.6 | 8.6 |
| IQ2_XS | 5.68 GiB | 6.6 | 6.6 | 6.7 | 7.0 | 7.5 | 8.5 |
| IQ2_XXS | 5.38 GiB | 6.3 | 6.3 | 6.4 | 6.7 | 7.2 | 8.2 |
| IQ1_M | 4.99 GiB | 5.9 | 5.9 | 6.1 | 6.3 | 6.8 | 7.8 |
| IQ1_S | 4.77 GiB | 5.6 | 5.7 | 5.8 | 6.1 | 6.6 | 7.6 |
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 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, this model uses latent attention and allocates no V cache at all, so any formula reading num_key_value_heads overstates its cache by more than an order of magnitude.