Which uses less memory at long context?

Qwen3.8-27B — its KV cache at 32K context is 5.0x smaller, because of how its attention layers cache.

Model comparison

Qwen3.8-27B vs DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-Reasoner

Q: Is Qwen3.8-27B or DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-Reasoner smaller?

Qwen3.8-27B is smaller at Q4_K_M: 15.66 GiB against 39.60 GiB.

At Q4_K_M, Qwen3.8-27B is the smaller download — 16,810,714,336 bytes against 42,520,406,496. At long context the gap widens: Qwen3.8-27B's KV cache at 32K is 5.0× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	Qwen3.8-27B	DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-Reasoner
Parameters	27.8B	70.6B
Architecture	qwen35	llama
Layers	64	80
Native context	262,144	131,072
Mixture of experts	no	no
Quantizations published	22	31
Smallest quantization	8.39 GiB	14.29 GiB
Q4_K_M	15.66 GiB	39.60 GiB
Licence	apache-2.0	llama3.3

KV cache by context

the term that decides long-context viability

Context	Qwen3.8-27B	DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-Reasoner	Ratio
4,096	0.25 GiB	1.25 GiB	5.00×
8,192	0.50 GiB	2.50 GiB	5.00×
16,384	1.00 GiB	5.00 GiB	5.00×
32,768	2.00 GiB	10.00 GiB	5.00×
65,536	4.00 GiB	20.00 GiB	5.00×
131,072	8.00 GiB	40.00 GiB	5.00×