Is VibeThinker-3B or Llama-3.1-8B-Instruct smaller?

VibeThinker-3B is smaller at Q4_K_M: 1.80 GiB against 4.58 GiB.

Which uses less memory at long context?

VibeThinker-3B — its KV cache at 32K context is 3.6x smaller, because of how its attention layers cache.

Model comparison

VibeThinker-3B vs Llama-3.1-8B-Instruct

At Q4_K_M, VibeThinker-3B is the smaller download — 1,929,902,176 bytes against 4,920,739,168. At long context the gap widens: VibeThinker-3B's KV cache at 32K is 3.6× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	VibeThinker-3B	Llama-3.1-8B-Instruct
Parameters	3.1B	8.0B
Architecture	qwen2	llama
Layers	36	32
Native context	131,072	131,072
Mixture of experts	no	no
Quantizations published	27	45
Smallest quantization	1.19 GiB	2.02 GiB
Q4_K_M	1.80 GiB	4.58 GiB
Licence	mit	llama3.1

KV cache by context

the term that decides long-context viability

Context	VibeThinker-3B	Llama-3.1-8B-Instruct	Ratio
4,096	0.14 GiB	0.50 GiB	3.56×
8,192	0.28 GiB	1.00 GiB	3.56×
16,384	0.56 GiB	2.00 GiB	3.56×
32,768	1.13 GiB	4.00 GiB	3.56×
65,536	2.25 GiB	8.00 GiB	3.56×
131,072	4.50 GiB	16.00 GiB	3.56×