Which uses less memory at long context?

Qwen3.6-27B — its KV cache at 32K context is 2.9x smaller, because of how its attention layers cache.

Model comparison

Qwen3.6-27B vs Llama3.2-30B-A3B-II-Dark-Champion-INSTRUCT-Heretic-Abliterated-Uncensored

Q: Is Qwen3.6-27B or Llama3.2-30B-A3B-II-Dark-Champion-INSTRUCT-Heretic-Abliterated-Uncensored smaller?

Qwen3.6-27B is smaller at Q4_K_M: 15.41 GiB against 17.11 GiB.

At Q4_K_M, Qwen3.6-27B is the smaller download — 16,547,397,888 bytes against 18,366,980,192. At long context the gap widens: Qwen3.6-27B's KV cache at 32K is 2.9× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	Qwen3.6-27B	Llama3.2-30B-A3B-II-Dark-Champion-INSTRUCT-Heretic-Abliterated-Uncensored
Parameters	27.8B	30.0B
Architecture	qwen35	llama
Layers	64	47
Native context	262,144	131,072
Mixture of experts	no	yes, 8 experts
Quantizations published	40	34
Smallest quantization	8.74 GiB	6.07 GiB
Q4_K_M	15.41 GiB	17.11 GiB
Licence	apache-2.0	apache-2.0

KV cache by context

the term that decides long-context viability

Context	Qwen3.6-27B	Llama3.2-30B-A3B-II-Dark-Champion-INSTRUCT-Heretic-Abliterated-Uncensored	Ratio
4,096	0.25 GiB	0.73 GiB	2.94×
8,192	0.50 GiB	1.47 GiB	2.94×
16,384	1.00 GiB	2.94 GiB	2.94×
32,768	2.00 GiB	5.88 GiB	2.94×
65,536	4.00 GiB	11.75 GiB	2.94×
131,072	8.00 GiB	23.50 GiB	2.94×