Model comparison
Qwen3.8-27B vs Llama-4-Scout-17B-16E-Instruct-4bit
At Q4_K_M, Qwen3.8-27B is the smaller download — 16,810,714,336 bytes against 22,491,854,144.
From the file· summed bytes, KV per layer
Side by side
| Qwen3.8-27B | Llama-4-Scout-17B-16E-Instruct-4bit | |
|---|---|---|
| Parameters | 27.8B | 17.0B |
| Architecture | qwen35 | llama |
| Layers | 64 | 48 |
| Native context | 262,144 | 10,485,760 |
| Mixture of experts | no | yes, 16 experts |
| Quantizations published | 22 | 3 |
| Smallest quantization | 8.39 GiB | 20.95 GiB |
| Q4_K_M | 15.66 GiB | 20.95 GiB |
| Licence | apache-2.0 | other |
KV cache by context
the term that decides long-context viability
| Context | Qwen3.8-27B | Llama-4-Scout-17B-16E-Instruct-4bit | Ratio |
|---|---|---|---|
| 4,096 | 0.25 GiB | — | — |
| 8,192 | 0.50 GiB | — | — |
| 16,384 | 1.00 GiB | — | — |
| 32,768 | 2.00 GiB | — | — |
| 65,536 | 4.00 GiB | — | — |
| 131,072 | 8.00 GiB | — | — |