Compare
Compare models side by side
Choose up to 4 models to see their file size, memory, license and third-party evaluations side by side.
Comparison table
| Item | gemma-4-12B-it-qat-UD-japanese-imatrix dahara1 ✕ Remove |
|---|---|
| Parameters | 11.9B |
| Architecture | gemma4 |
| Max context length | 262,144 |
| Representative quantization and size Measured | Q4_K_M · 7.12 GB |
| Memory at 8K Estimated | 8.83 GB |
| Memory at 32K Estimated | 9.24 GB |
| VRAM it fits (8K) Estimated | VRAM 12 GB+ |
| Smallest quantization (8K) Estimated | UD-IQ2_M · 4.21 GB → 5.92 GB |
| License | Apache License 2.0 Commercial OK |
| llama.cpp support | llama.cpp b8637 and later |
| Downloads (30 days, total) | 2,086 |
| Human preference (LMArena) | - |
| Capabilities (Epoch AI) | - |
| Hard science (GPQA) | - |
- The representative quantization is Q4_K_M (or the smallest one if missing) from the source with the most downloads. Memory is estimated with an f16 KV cache.
- Third-party evaluations are other organizations' values for the original model before quantization, not this site's rating (sources are on the ranking page).
- This site does not judge which model is better or give an overall score.