gemma-4-12b-it - DASH-Q 2-bit GGUF
llama.cpp GGUF files of google/gemma-4-12B-it quantized with DASH-Q, at several 2-bit-class sizes. Every file uses only standard llama.cpp tensor types (no tensor above 4 bits) and loads in any recent llama.cpp build.
| File | Type | Size | Bits / weight |
|---|---|---|---|
gemma-4-12b-it-DASHQ-IQ2_XXS.gguf |
IQ2_XXS | 3.65 GB | 2.45 |
gemma-4-12b-it-DASHQ-IQ2_M.gguf |
IQ2_M | 4.35 GB | 2.93 |
gemma-4-12b-it-DASHQ-Q2_K_XL.gguf |
Q2_K_XL | 4.67 GB | 3.14 |
Text only: the vision tower is not included.
Quality vs. the unquantized model (KL divergence)
Raw-text perplexity is not a meaningful quality metric for this instruction/reasoning model (the unquantized model itself scores
poorly on WikiText-2 while generating normally), so quality is reported as KL divergence of the next-token distribution to the
unquantized reference (llama-perplexity --kl-divergence, WikiText-2, 4 x 2048 tokens):
| Type | Model | Size | Mean KLD | Median KLD | Top-1 token agreement |
|---|---|---|---|---|---|
| IQ2_XXS | llama.cpp IQ2_XXS (imatrix) | 3.55 GB | 9.86 | 9.53 | 2.2% |
| IQ2_XXS | DASH-Q IQ2_XXS | 3.65 GB | 3.56 | 3.04 | 25.3% |
| IQ2_XS | llama.cpp IQ2_XS (imatrix) | 3.87 GB | 6.35 | 5.91 | 10.4% |
| IQ2_M | llama.cpp IQ2_M (imatrix) | 4.36 GB | 4.81 | 4.39 | 15.0% |
| IQ2_M | unsloth UD-IQ2_M | 4.21 GB | 4.56 | 4.15 | 22.5% |
| IQ2_M | DASH-Q IQ2_M | 4.35 GB | 1.90 | 1.19 | 47.8% |
| Q2_K_XL | llama.cpp Q2_K (imatrix) | 4.81 GB | 3.80 | 3.40 | 24.3% |
| Q2_K_XL | unsloth UD-Q2_K_XL | 4.66 GB | 3.41 | 2.94 | 32.0% |
| Q2_K_XL | DASH-Q Q2_K_XL | 4.67 GB | 1.94 | 1.16 | 48.6% |
Usage
llama-cli -m gemma-4-12b-it-DASHQ-Q2_K_XL.gguf -ngl 99 -c 8192
License
Inherits the license of the base model (google/gemma-4-12B-it).