Model comparison
ERNIE-21B-A3B-Thinking-Gemini-3-Pro-High-Reasoning-V2 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: ERNIE-21B-A3B-Thinking-Gemini-3-Pro-High-Reasoning-V2's KV cache at 32K is 1.4× smaller, which usually matters more than the difference in weights.
From the file· summed bytes, KV per layer
Side by side
| ERNIE-21B-A3B-Thinking-Gemini-3-Pro-High-Reasoning-V2 | gemma-4-12B-it-qat-q4_0-unquantized | |
|---|---|---|
| Parameters | 21.8B | 12.0B |
| Architecture | ernie4_5-moe | gemma4 |
| Layers | 28 | 48 |
| Native context | 131,072 | 262,144 |
| Mixture of experts | no | no |
| Quantizations published | 23 | 2 |
| Smallest quantization | 4.17 GiB | 6.50 GiB |
| Q4_K_M | — | — |
| Licence | apache-2.0 | apache-2.0 |
KV cache by context
the term that decides long-context viability
| Context | ERNIE-21B-A3B-Thinking-Gemini-3-Pro-High-Reasoning-V2 | gemma-4-12B-it-qat-q4_0-unquantized | Ratio |
|---|---|---|---|
| 4,096 | 0.22 GiB | 0.72 GiB | 3.29× |
| 8,192 | 0.44 GiB | 0.97 GiB | 2.21× |
| 16,384 | 0.88 GiB | 1.47 GiB | 1.68× |
| 32,768 | 1.75 GiB | 2.47 GiB | 1.41× |
| 65,536 | 3.50 GiB | 4.47 GiB | 1.28× |
| 131,072 | 7.00 GiB | 8.47 GiB | 1.21× |