Model comparison
GPT-NeoX-20B-Erebus vs Qwen3.8-27B
These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: Qwen3.8-27B's KV cache at 32K is 16.5× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| GPT-NeoX-20B-Erebus | Qwen3.8-27B | |
|---|---|---|
| Parameters | 20.6B | 27.8B |
| Architecture | gptneox | qwen35 |
| Layers | 44 | 64 |
| Native context | 2,048 | 262,144 |
| Mixture of experts | no | no |
| Quantizations published | 23 | 22 |
| Smallest quantization | 4.12 GiB | 8.39 GiB |
| Q4_K_M | — | 15.66 GiB |
| Licence | apache-2.0 | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | GPT-NeoX-20B-Erebus | Qwen3.8-27B | Ratio |
|---|---|---|---|
| 4,096 | 4.13 GiB | 0.25 GiB | 16.49× |
| 8,192 | 8.25 GiB | 0.50 GiB | 16.49× |
| 16,384 | 16.50 GiB | 1.00 GiB | 16.50× |
| 32,768 | 33.00 GiB | 2.00 GiB | 16.50× |
| 65,536 | 66.00 GiB | 4.00 GiB | 16.50× |
| 131,072 | 132.00 GiB | 8.00 GiB | 16.50× |