Model comparison

DeepSeek-Coder-V2-Lite-Base vs gemma-4-12B-it-qat-q4_0-unquantized

These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: DeepSeek-Coder-V2-Lite-Base's KV cache at 32K is 2.6× smaller, which usually matters more than the difference in weights.

From the file· summed bytes, KV per layer

Side by side

DeepSeek-Coder-V2-Lite-Basegemma-4-12B-it-qat-q4_0-unquantized
Parameters15.7B12.0B
Architecturedeepseek2gemma4
Layers2748
Native context163,840262,144
Mixture of expertsyes, 64 expertsno
Quantizations published462
Smallest quantization4.65 GiB6.50 GiB
Q4_K_M9.65 GiB—
Licenceotherapache-2.0

KV cache by context

the term that decides long-context viability
ContextDeepSeek-Coder-V2-Lite-Basegemma-4-12B-it-qat-q4_0-unquantizedRatio
4,0960.12 GiB0.72 GiB6.06×
8,1920.24 GiB0.97 GiB4.08×
16,3840.47 GiB1.47 GiB3.09×
32,7680.95 GiB2.47 GiB2.60×
65,5361.90 GiB4.47 GiB2.35×
131,0723.80 GiB8.47 GiB2.23×
DeepSeek-Coder-V2-Lite-Base vs gemma-4-12B-it-qat-q4_0-unquantized — size, memory and hardware fit — ossmodeldb