Is Qwen3-VL-4B-Thinking-Unredacted-MAX or Llama-3.1-8B-Instruct smaller?

Qwen3-VL-4B-Thinking-Unredacted-MAX is smaller at Q4_K_M: 2.33 GiB against 4.58 GiB.

Which uses less memory at long context?

Llama-3.1-8B-Instruct — its KV cache at 32K context is 1.1x smaller, because of how its attention layers cache.

Model comparison

At Q4_K_M, Qwen3-VL-4B-Thinking-Unredacted-MAX is the smaller download — 2,497,282,656 bytes against 4,920,739,168.

From the file· summed bytes, KV per layer

	Qwen3-VL-4B-Thinking-Unredacted-MAX	Llama-3.1-8B-Instruct
Parameters	4.4B	8.0B
Architecture	qwen3vl	llama
Layers	36	32
Native context	262,144	131,072
Mixture of experts	no	no
Quantizations published	36	45
Smallest quantization	0.98 GiB	2.02 GiB
Q4_K_M	2.33 GiB	4.58 GiB
Licence	apache-2.0	llama3.1

the term that decides long-context viability

Context	Qwen3-VL-4B-Thinking-Unredacted-MAX	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×