Is Qwen3-VL-4B-Thinking or Qwen2.5-VL-7B-Instruct smaller?

Qwen3-VL-4B-Thinking is smaller at Q4_K_M: 2.33 GiB against 4.36 GiB.

Which uses less memory at long context?

Qwen2.5-VL-7B-Instruct — its KV cache at 32K context is 2.6x smaller, because of how its attention layers cache.

Model comparison

Qwen3-VL-4B-Thinking vs Qwen2.5-VL-7B-Instruct

At Q4_K_M, Qwen3-VL-4B-Thinking is the smaller download — 2,497,281,472 bytes against 4,683,072,032. At long context the gap widens: Qwen2.5-VL-7B-Instruct's KV cache at 32K is 2.6× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	Qwen3-VL-4B-Thinking	Qwen2.5-VL-7B-Instruct
Parameters	4.4B	8.3B
Architecture	qwen3vl	qwen2vl
Layers	36	28
Native context	262,144	128,000
Mixture of experts	no	no
Quantizations published	43	50
Smallest quantization	1.01 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-Thinking	Qwen2.5-VL-7B-Instruct	Ratio
4,096	0.56 GiB	0.22 GiB	2.57×
8,192	1.13 GiB	0.44 GiB	2.57×
16,384	2.25 GiB	0.88 GiB	2.57×
32,768	4.50 GiB	1.75 GiB	2.57×
65,536	9.00 GiB	3.50 GiB	2.57×
131,072	18.00 GiB	7.00 GiB	2.57×