Model comparison

Qwen3-Coder-30B-A3B-Instruct vs ERNIE-21B-A3B-Thinking-Gemini-3-Pro-High-Reasoning-V2

These two publish different quantization sets; the table below has the exact sizes. At long context the gap widens: ERNIE-21B-A3B-Thinking-Gemini-3-Pro-High-Reasoning-V2'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

Qwen3-Coder-30B-A3B-InstructERNIE-21B-A3B-Thinking-Gemini-3-Pro-High-Reasoning-V2
Parameters30.5B21.8B
Architectureqwen3moeernie4_5-moe
Layers4828
Native context262,144131,072
Mixture of expertsyes, 128 expertsno
Quantizations published4623
Smallest quantization7.46 GiB4.17 GiB
Q4_K_M17.28 GiB—
Licenceapache-2.0apache-2.0

KV cache by context

the term that decides long-context viability
ContextQwen3-Coder-30B-A3B-InstructERNIE-21B-A3B-Thinking-Gemini-3-Pro-High-Reasoning-V2Ratio
4,0960.38 GiB0.22 GiB1.71×
8,1920.75 GiB0.44 GiB1.71×
16,3841.50 GiB0.88 GiB1.71×
32,7683.00 GiB1.75 GiB1.71×
65,5366.00 GiB3.50 GiB1.71×
131,07212.00 GiB7.00 GiB1.71×
Qwen3-Coder-30B-A3B-Instruct vs ERNIE-21B-A3B-Thinking-Gemini-3-Pro-High-Reasoning-V2 — size, memory and hardware fit — ossmodeldb