Is Holo-3.1-0.8B or Llama-3.2-3B-Instruct smaller?

Holo-3.1-0.8B is smaller at Q4_K_M: 0.63 GiB against 1.88 GiB.

Which uses less memory at long context?

Holo-3.1-0.8B — its KV cache at 32K context is 9.3x smaller, because of how its attention layers cache.

Model comparison

Holo-3.1-0.8B vs Llama-3.2-3B-Instruct

At Q4_K_M, Holo-3.1-0.8B is the smaller download — 672,330,784 bytes against 2,019,377,440. At long context the gap widens: Holo-3.1-0.8B's KV cache at 32K is 9.3× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	Holo-3.1-0.8B	Llama-3.2-3B-Instruct
Parameters	1.1B	3.2B
Architecture	qwen35	llama
Layers	24	28
Native context	262,144	131,072
Mixture of experts	no	no
Quantizations published	36	39
Smallest quantization	0.38 GiB	0.85 GiB
Q4_K_M	0.63 GiB	1.88 GiB
Licence	apache-2.0	—

KV cache by context

the term that decides long-context viability

Context	Holo-3.1-0.8B	Llama-3.2-3B-Instruct	Ratio
4,096	0.05 GiB	0.44 GiB	9.33×
8,192	0.09 GiB	0.88 GiB	9.33×
16,384	0.19 GiB	1.75 GiB	9.33×
32,768	0.38 GiB	3.50 GiB	9.33×
65,536	0.75 GiB	7.00 GiB	9.33×
131,072	1.50 GiB	14.00 GiB	9.33×