Model comparison

smollm-360M-instruct-add-basics vs MiniCPM5-1B-Claude-Opus-Fable5-Thinking

These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: MiniCPM5-1B-Claude-Opus-Fable5-Thinking's KV cache at 32K is 1.7× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

smollm-360M-instruct-add-basicsMiniCPM5-1B-Claude-Opus-Fable5-Thinking
Parameters362M1.1B
Architecturellamallama
Layers3224
Native context2,048131,072
Mixture of expertsnono
Quantizations published815
Smallest quantization0.19 GiB0.43 GiB
Q4_K_M0.64 GiB
Licenceapache-2.0

KV cache by context

the term that decides long-context viability
Contextsmollm-360M-instruct-add-basicsMiniCPM5-1B-Claude-Opus-Fable5-ThinkingRatio
4,0960.16 GiB0.09 GiB1.67×
8,1920.31 GiB0.19 GiB1.67×
16,3840.63 GiB0.38 GiB1.67×
32,7681.25 GiB0.75 GiB1.67×
65,5362.50 GiB1.50 GiB1.67×
131,0725.00 GiB3.00 GiB1.67×