Model comparison
DeepSeek-Coder-V2-Instruct-0724 vs Hy3
At Q4_K_M, DeepSeek-Coder-V2-Instruct-0724 is the smaller download — 142,453,966,944 bytes against 180,999,059,328. At long context the gap widens: DeepSeek-Coder-V2-Instruct-0724'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-Coder-V2-Instruct-0724 | Hy3 | |
|---|---|---|
| Parameters | 236B | 299B |
| Architecture | deepseek2 | hy_v3 |
| Layers | 60 | 80 |
| Native context | 163,840 | 262,144 |
| Mixture of experts | yes, 160 experts | yes, 192 experts |
| Quantizations published | 44 | 37 |
| Smallest quantization | 44.14 GiB | 59.65 GiB |
| Q4_K_M | 132.67 GiB | 168.57 GiB |
| Licence | other | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | DeepSeek-Coder-V2-Instruct-0724 | Hy3 | Ratio |
|---|---|---|---|
| 4,096 | 0.26 GiB | 1.25 GiB | 4.74× |
| 8,192 | 0.53 GiB | 2.50 GiB | 4.74× |
| 16,384 | 1.05 GiB | 5.00 GiB | 4.74× |
| 32,768 | 2.11 GiB | 10.00 GiB | 4.74× |
| 65,536 | 4.22 GiB | 20.00 GiB | 4.74× |
| 131,072 | 8.44 GiB | 40.00 GiB | 4.74× |