Model comparison
DeepSeek-V3.1 vs Hy3
At Q4_K_M, Hy3 is the smaller download — 180,999,059,328 bytes against 404,494,158,304. At long context the gap widens: DeepSeek-V3.1's KV cache at 32K is 4.7× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| DeepSeek-V3.1 | Hy3 | |
|---|---|---|
| Parameters | 685B | 299B |
| Architecture | deepseek2 | hy_v3 |
| Layers | 61 | 80 |
| Native context | 163,840 | 262,144 |
| Mixture of experts | yes, 256 experts | yes, 192 experts |
| Quantizations published | 50 | 37 |
| Smallest quantization | 137.32 GiB | 59.65 GiB |
| Q4_K_M | 376.71 GiB | 168.57 GiB |
| Licence | mit | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | DeepSeek-V3.1 | Hy3 | Ratio |
|---|---|---|---|
| 4,096 | 0.27 GiB | 1.25 GiB | 4.66× |
| 8,192 | 0.54 GiB | 2.50 GiB | 4.66× |
| 16,384 | 1.07 GiB | 5.00 GiB | 4.66× |
| 32,768 | 2.14 GiB | 10.00 GiB | 4.66× |
| 65,536 | 4.29 GiB | 20.00 GiB | 4.66× |
| 131,072 | 8.58 GiB | 40.00 GiB | 4.66× |