Model comparison

HyperCLOVAX-SEED-Text-Instruct-0.5B vs Llama-3.2-3B-Instruct

At Q4_K_M, HyperCLOVAX-SEED-Text-Instruct-0.5B is the smaller download — 431,882,784 bytes against 2,019,377,440. At long context the gap widens: HyperCLOVAX-SEED-Text-Instruct-0.5B'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

HyperCLOVAX-SEED-Text-Instruct-0.5BLlama-3.2-3B-Instruct
Parameters566M3.2B
Architecturellamallama
Layers2428
Native context8,192131,072
Mixture of expertsnono
Quantizations published439
Smallest quantization0.39 GiB0.85 GiB
Q4_K_M0.40 GiB1.88 GiB
Licenceother

KV cache by context

the term that decides long-context viability
ContextHyperCLOVAX-SEED-Text-Instruct-0.5BLlama-3.2-3B-InstructRatio
4,0960.38 GiB0.44 GiB1.17×
8,1920.75 GiB0.88 GiB1.17×
16,3841.50 GiB1.75 GiB1.17×
32,7683.00 GiB3.50 GiB1.17×
65,5366.00 GiB7.00 GiB1.17×
131,07212.00 GiB14.00 GiB1.17×