Model comparison
KAT-Coder-V2.5-Dev vs gemma-4-12B-it-qat-q4_0-unquantized
These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: KAT-Coder-V2.5-Dev's KV cache at 32K is 3.9× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| KAT-Coder-V2.5-Dev | gemma-4-12B-it-qat-q4_0-unquantized | |
|---|---|---|
| Parameters | 34.7B | 12.0B |
| Architecture | qwen35moe | gemma4 |
| Layers | 40 | 48 |
| Native context | 262,144 | 262,144 |
| Mixture of experts | yes, 256 experts | no |
| Quantizations published | 29 | 2 |
| Smallest quantization | 9.11 GiB | 6.50 GiB |
| Q4_K_M | 19.92 GiB | — |
| Licence | apache-2.0 | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | KAT-Coder-V2.5-Dev | gemma-4-12B-it-qat-q4_0-unquantized | Ratio |
|---|---|---|---|
| 4,096 | 0.08 GiB | 0.72 GiB | 9.19× |
| 8,192 | 0.16 GiB | 0.97 GiB | 6.20× |
| 16,384 | 0.31 GiB | 1.47 GiB | 4.70× |
| 32,768 | 0.63 GiB | 2.47 GiB | 3.95× |
| 65,536 | 1.25 GiB | 4.47 GiB | 3.57× |
| 131,072 | 2.50 GiB | 8.47 GiB | 3.39× |