Model comparison
Marco-Nano-Instruct vs Qwen3-8B
At Q4_K_M, Qwen3-8B is the smaller download — 5,027,783,488 bytes against 5,371,989,536. At long context the gap widens: Marco-Nano-Instruct's KV cache at 32K is 1.3× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| Marco-Nano-Instruct | Qwen3-8B | |
|---|---|---|
| Parameters | 8.0B | 8.2B |
| Architecture | qwen3moe | qwen3 |
| Layers | 28 | 36 |
| Native context | 32,768 | 40,960 |
| Mixture of experts | yes, 232 experts | no |
| Quantizations published | 36 | 51 |
| Smallest quantization | 2.44 GiB | 2.12 GiB |
| Q4_K_M | 5.00 GiB | 4.68 GiB |
| Licence | apache-2.0 | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | Marco-Nano-Instruct | Qwen3-8B | Ratio |
|---|---|---|---|
| 4,096 | 0.44 GiB | 0.56 GiB | 1.29× |
| 8,192 | 0.88 GiB | 1.13 GiB | 1.29× |
| 16,384 | 1.75 GiB | 2.25 GiB | 1.29× |
| 32,768 | 3.50 GiB | 4.50 GiB | 1.29× |
| 65,536 | 7.00 GiB | 9.00 GiB | 1.29× |
| 131,072 | 14.00 GiB | 18.00 GiB | 1.29× |