Model comparison
Qwen3-4B-Thinking-2507 vs Llama-3.1-8B-Instruct
At Q4_K_M, Qwen3-4B-Thinking-2507 is the smaller download — 2,497,280,448 bytes against 4,920,739,168.
From the file· summed bytes, KV per layer
Side by side
| Qwen3-4B-Thinking-2507 | Llama-3.1-8B-Instruct | |
|---|---|---|
| Parameters | 4.0B | 8.0B |
| Architecture | qwen3 | llama |
| Layers | 36 | 32 |
| Native context | 262,144 | 131,072 |
| Mixture of experts | no | no |
| Quantizations published | 47 | 45 |
| Smallest quantization | 1.01 GiB | 2.02 GiB |
| Q4_K_M | 2.33 GiB | 4.58 GiB |
| Licence | — | llama3.1 |
KV cache by context
the term that decides long-context viability
| Context | Qwen3-4B-Thinking-2507 | Llama-3.1-8B-Instruct | Ratio |
|---|---|---|---|
| 4,096 | 0.56 GiB | 0.50 GiB | 1.13× |
| 8,192 | 1.13 GiB | 1.00 GiB | 1.13× |
| 16,384 | 2.25 GiB | 2.00 GiB | 1.13× |
| 32,768 | 4.50 GiB | 4.00 GiB | 1.13× |
| 65,536 | 9.00 GiB | 8.00 GiB | 1.13× |
| 131,072 | 18.00 GiB | 16.00 GiB | 1.13× |