Is bitnet_b1_58-large or Llama-3.2-1B-Instruct smaller?

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

Which uses less memory at long context?

Llama-3.2-1B-Instruct — its KV cache at 32K context is 4.5x smaller, because of how its attention layers cache.

Model comparison

bitnet_b1_58-large vs Llama-3.2-1B-Instruct

These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: Llama-3.2-1B-Instruct's KV cache at 32K is 4.5× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	bitnet_b1_58-large	Llama-3.2-1B-Instruct
Parameters	729M	1.2B
Architecture	bitnet	llama
Layers	24	16
Native context	2,048	131,072
Mixture of experts	no	no
Quantizations published	1	39
Smallest quantization	0.20 GiB	0.39 GiB
Q4_K_M	—	0.75 GiB
Licence	mit	—

KV cache by context

the term that decides long-context viability

Context	bitnet_b1_58-large	Llama-3.2-1B-Instruct	Ratio
4,096	0.56 GiB	0.13 GiB	4.50×
8,192	1.13 GiB	0.25 GiB	4.50×
16,384	2.25 GiB	0.50 GiB	4.50×
32,768	4.50 GiB	1.00 GiB	4.50×
65,536	9.00 GiB	2.00 GiB	4.50×
131,072	18.00 GiB	4.00 GiB	4.50×