Model comparison

reasoning-summarizer-800m-pre vs Llama-3.2-1B-Instruct

At Q4_K_M, reasoning-summarizer-800m-pre is the smaller download — 529,289,792 bytes against 807,690,688. At long context the gap widens: reasoning-summarizer-800m-pre's KV cache at 32K is 2.7× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

reasoning-summarizer-800m-preLlama-3.2-1B-Instruct
Parameters752M1.2B
Architectureqwen35llama
Layers2416
Native context262,144131,072
Mixture of expertsnono
Quantizations published639
Smallest quantization0.47 GiB0.39 GiB
Q4_K_M0.49 GiB0.75 GiB
Licenceapache-2.0—

KV cache by context

the term that decides long-context viability
Contextreasoning-summarizer-800m-preLlama-3.2-1B-InstructRatio
4,0960.05 GiB0.13 GiB2.66×
8,1920.09 GiB0.25 GiB2.67×
16,3840.19 GiB0.50 GiB2.67×
32,7680.38 GiB1.00 GiB2.67×
65,5360.75 GiB2.00 GiB2.67×
131,0721.50 GiB4.00 GiB2.67×