Is deepseek-llm-67b-base or Qwen3-30B-A3B-Thinking-2507 smaller?

Qwen3-30B-A3B-Thinking-2507 is smaller at Q4_K_M: 17.28 GiB against 37.66 GiB.

Which uses less memory at long context?

Qwen3-30B-A3B-Thinking-2507 — its KV cache at 32K context is 4.0x smaller, because of how its attention layers cache.

Model comparison

deepseek-llm-67b-base vs Qwen3-30B-A3B-Thinking-2507

At Q4_K_M, Qwen3-30B-A3B-Thinking-2507 is the smaller download — 18,556,685,824 bytes against 40,439,686,592. At long context the gap widens: Qwen3-30B-A3B-Thinking-2507's KV cache at 32K is 4.0× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	deepseek-llm-67b-base	Qwen3-30B-A3B-Thinking-2507
Parameters	67.4B	30.5B
Architecture	llama	qwen3moe
Layers	95	48
Native context	4,096	262,144
Mixture of experts	no	yes, 128 experts
Quantizations published	30	51
Smallest quantization	13.65 GiB	7.05 GiB
Q4_K_M	37.66 GiB	17.28 GiB
Licence	other	apache-2.0

KV cache by context

the term that decides long-context viability

Context	deepseek-llm-67b-base	Qwen3-30B-A3B-Thinking-2507	Ratio
4,096	1.48 GiB	0.38 GiB	3.96×
8,192	2.97 GiB	0.75 GiB	3.96×
16,384	5.94 GiB	1.50 GiB	3.96×
32,768	11.88 GiB	3.00 GiB	3.96×
65,536	23.75 GiB	6.00 GiB	3.96×
131,072	47.50 GiB	12.00 GiB	3.96×