Model comparison
gemma-4-12B-it-qat-q4_0-unquantized vs Moonlight-16B-A3B-Instruct
These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: Moonlight-16B-A3B-Instruct's KV cache at 32K is 2.6× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| gemma-4-12B-it-qat-q4_0-unquantized | Moonlight-16B-A3B-Instruct | |
|---|---|---|
| Parameters | 12.0B | 16.0B |
| Architecture | gemma4 | deepseek2 |
| Layers | 48 | 27 |
| Native context | 262,144 | 8,192 |
| Mixture of experts | no | yes, 64 experts |
| Quantizations published | 2 | 48 |
| Smallest quantization | 6.50 GiB | 4.77 GiB |
| Q4_K_M | — | 9.81 GiB |
| Licence | apache-2.0 | mit |
KV cache by context
the term that decides long-context viability
| Context | gemma-4-12B-it-qat-q4_0-unquantized | Moonlight-16B-A3B-Instruct | Ratio |
|---|---|---|---|
| 4,096 | 0.72 GiB | 0.12 GiB | 6.06× |
| 8,192 | 0.97 GiB | 0.24 GiB | 4.08× |
| 16,384 | 1.47 GiB | 0.47 GiB | 3.09× |
| 32,768 | 2.47 GiB | 0.95 GiB | 2.60× |
| 65,536 | 4.47 GiB | 1.90 GiB | 2.35× |
| 131,072 | 8.47 GiB | 3.80 GiB | 2.23× |