Is Phi-3-medium-4k-instruct or llama-3-youko-8b smaller?

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

Which uses less memory at long context?

llama-3-youko-8b — its KV cache at 32K context is 1.6x smaller, because of how its attention layers cache.

Model comparison

Phi-3-medium-4k-instruct vs llama-3-youko-8b

These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: llama-3-youko-8b's KV cache at 32K is 1.6× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

	Phi-3-medium-4k-instruct	llama-3-youko-8b
Parameters	14.0B	8.0B
Architecture	phi3	llama
Layers	40	32
Native context	4,096	8,192
Mixture of experts	no	no
Quantizations published	54	2
Smallest quantization	2.75 GiB	5.34 GiB
Q4_K_M	7.98 GiB	—
Licence	mit	llama3

KV cache by context

the term that decides long-context viability

Context	Phi-3-medium-4k-instruct	llama-3-youko-8b	Ratio
4,096	0.78 GiB	0.50 GiB	1.56×
8,192	1.56 GiB	1.00 GiB	1.56×
16,384	3.13 GiB	2.00 GiB	1.56×
32,768	6.25 GiB	4.00 GiB	1.56×
65,536	12.50 GiB	8.00 GiB	1.56×
131,072	25.00 GiB	16.00 GiB	1.56×