Model comparison
Llama-3_1-Nemotron-51B-Instruct vs Qwen3-Coder-30B-A3B-Instruct
At Q4_K_M, Qwen3-Coder-30B-A3B-Instruct is the smaller download — 18,556,689,568 bytes against 31,037,307,136. At long context the gap widens: Qwen3-Coder-30B-A3B-Instruct's KV cache at 32K is 26.7× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| Llama-3_1-Nemotron-51B-Instruct | Qwen3-Coder-30B-A3B-Instruct | |
|---|---|---|
| Parameters | 51.5B | 30.5B |
| Architecture | deci | qwen3moe |
| Layers | 80 | 48 |
| Native context | 131,072 | 262,144 |
| Mixture of experts | no | yes, 128 experts |
| Quantizations published | 25 | 46 |
| Smallest quantization | 13.07 GiB | 7.46 GiB |
| Q4_K_M | 28.91 GiB | 17.28 GiB |
| Licence | other | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | Llama-3_1-Nemotron-51B-Instruct | Qwen3-Coder-30B-A3B-Instruct | Ratio |
|---|---|---|---|
| 4,096 | 10.00 GiB | 0.38 GiB | 26.67× |
| 8,192 | 20.00 GiB | 0.75 GiB | 26.67× |
| 16,384 | 40.00 GiB | 1.50 GiB | 26.67× |
| 32,768 | 80.00 GiB | 3.00 GiB | 26.67× |
| 65,536 | 160.00 GiB | 6.00 GiB | 26.67× |
| 131,072 | 320.00 GiB | 12.00 GiB | 26.67× |