Is Felldude-Uncensored-Ministral3-3B-bf16 or Llama-3.1-8B-Instruct smaller?

Both publish quantizations; compare the table for the exact sizes.

Which uses less memory at long context?

Felldude-Uncensored-Ministral3-3B-bf16 — its KV cache at 32K context is 1.2x smaller, because of how its attention layers cache.

Model comparison

Felldude-Uncensored-Ministral3-3B-bf16 vs Llama-3.1-8B-Instruct

These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: Felldude-Uncensored-Ministral3-3B-bf16's KV cache at 32K is 1.2× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	Felldude-Uncensored-Ministral3-3B-bf16	Llama-3.1-8B-Instruct
Parameters	3.8B	8.0B
Architecture	mistral3	llama
Layers	26	32
Native context	262,144	131,072
Mixture of experts	no	no
Quantizations published	24	45
Smallest quantization	0.87 GiB	2.02 GiB
Q4_K_M	—	4.58 GiB
Licence	—	llama3.1

KV cache by context

the term that decides long-context viability

Context	Felldude-Uncensored-Ministral3-3B-bf16	Llama-3.1-8B-Instruct	Ratio
4,096	0.41 GiB	0.50 GiB	1.23×
8,192	0.81 GiB	1.00 GiB	1.23×
16,384	1.63 GiB	2.00 GiB	1.23×
32,768	3.25 GiB	4.00 GiB	1.23×
65,536	6.50 GiB	8.00 GiB	1.23×
131,072	13.00 GiB	16.00 GiB	1.23×