Model comparison
Qwen3-Coder-30B-A3B-Instruct vs GPT-NeoX-20B-Erebus
These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: Qwen3-Coder-30B-A3B-Instruct's KV cache at 32K is 11.0× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| Qwen3-Coder-30B-A3B-Instruct | GPT-NeoX-20B-Erebus | |
|---|---|---|
| Parameters | 30.5B | 20.6B |
| Architecture | qwen3moe | gptneox |
| Layers | 48 | 44 |
| Native context | 262,144 | 2,048 |
| Mixture of experts | yes, 128 experts | no |
| Quantizations published | 46 | 23 |
| Smallest quantization | 7.46 GiB | 4.12 GiB |
| Q4_K_M | 17.28 GiB | — |
| Licence | apache-2.0 | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | Qwen3-Coder-30B-A3B-Instruct | GPT-NeoX-20B-Erebus | Ratio |
|---|---|---|---|
| 4,096 | 0.38 GiB | 4.13 GiB | 11.00× |
| 8,192 | 0.75 GiB | 8.25 GiB | 11.00× |
| 16,384 | 1.50 GiB | 16.50 GiB | 11.00× |
| 32,768 | 3.00 GiB | 33.00 GiB | 11.00× |
| 65,536 | 6.00 GiB | 66.00 GiB | 11.00× |
| 131,072 | 12.00 GiB | 132.00 GiB | 11.00× |