Model comparison
Qwen3-VL-4B-Thinking vs Qwen3-VL-4B-Instruct
At Q4_K_M, Qwen3-VL-4B-Thinking is the smaller download — 2,497,281,472 bytes against 2,497,281,568.
From the file· summed bytes, KV per layer
Side by side
| Qwen3-VL-4B-Thinking | Qwen3-VL-4B-Instruct | |
|---|---|---|
| Parameters | 4.4B | 4.4B |
| Architecture | qwen3vl | qwen3vl |
| Layers | 36 | 36 |
| Native context | 262,144 | 262,144 |
| Mixture of experts | no | no |
| Quantizations published | 43 | 26 |
| Smallest quantization | 1.01 GiB | 1.01 GiB |
| Q4_K_M | 2.33 GiB | 2.33 GiB |
| Licence | apache-2.0 | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | Qwen3-VL-4B-Thinking | Qwen3-VL-4B-Instruct | Ratio |
|---|---|---|---|
| 4,096 | 0.56 GiB | 0.56 GiB | 1.00× |
| 8,192 | 1.13 GiB | 1.13 GiB | 1.00× |
| 16,384 | 2.25 GiB | 2.25 GiB | 1.00× |
| 32,768 | 4.50 GiB | 4.50 GiB | 1.00× |
| 65,536 | 9.00 GiB | 9.00 GiB | 1.00× |
| 131,072 | 18.00 GiB | 18.00 GiB | 1.00× |