Model comparison

MARTHA-LXVII.8B_QWEN-3.5-3.6_prune_9b-3.6_base vs llama-3-youko-8b

These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: MARTHA-LXVII.8B_QWEN-3.5-3.6_prune_9b-3.6_base's KV cache at 32K is 4.6× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

MARTHA-LXVII.8B_QWEN-3.5-3.6_prune_9b-3.6_basellama-3-youko-8b
Parameters8.1B8.0B
Architectureqwen35llama
Layers2832
Native context262,1448,192
Mixture of expertsnono
Quantizations published362
Smallest quantization2.35 GiB5.34 GiB
Q4_K_M4.76 GiB
Licenceapache-2.0llama3

KV cache by context

the term that decides long-context viability
ContextMARTHA-LXVII.8B_QWEN-3.5-3.6_prune_9b-3.6_basellama-3-youko-8bRatio
4,0960.11 GiB0.50 GiB4.57×
8,1920.22 GiB1.00 GiB4.57×
16,3840.44 GiB2.00 GiB4.57×
32,7680.88 GiB4.00 GiB4.57×
65,5361.75 GiB8.00 GiB4.57×
131,0723.50 GiB16.00 GiB4.57×