Model comparison
Qwen3-VL-4B-Instruct-abliterated-v1 vs Qwen2.5-VL-7B-Instruct
At Q4_K_M, Qwen3-VL-4B-Instruct-abliterated-v1 is the smaller download — 2,497,282,560 bytes against 4,683,072,032.
From the file· summed bytes, KV per layer
Side by side
| Qwen3-VL-4B-Instruct-abliterated-v1 | Qwen2.5-VL-7B-Instruct | |
|---|---|---|
| Parameters | 4.4B | 8.3B |
| Architecture | qwen3vl | qwen2vl |
| Layers | — | 28 |
| Native context | — | 128,000 |
| Mixture of experts | no | no |
| Quantizations published | 60 | 50 |
| Smallest quantization | 0.98 GiB | 1.93 GiB |
| Q4_K_M | 2.33 GiB | 4.36 GiB |
| Licence | apache-2.0 | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | Qwen3-VL-4B-Instruct-abliterated-v1 | Qwen2.5-VL-7B-Instruct | Ratio |
|---|---|---|---|
| 4,096 | — | 0.22 GiB | — |
| 8,192 | — | 0.44 GiB | — |
| 16,384 | — | 0.88 GiB | — |
| 32,768 | — | 1.75 GiB | — |
| 65,536 | — | 3.50 GiB | — |
| 131,072 | — | 7.00 GiB | — |