gemma-4-31B-it - DASH-Q 2-bit GGUF
llama.cpp GGUF files of google/gemma-4-31B-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-4-31B-it-DASHQ-IQ2_XXS.gguf |
IQ2_XXS | 8.82 GB | 2.30 |
gemma-4-31B-it-DASHQ-IQ2_XS.gguf |
IQ2_XS | 10.03 GB | 2.61 |
gemma-4-31B-it-DASHQ-IQ2_M.gguf |
IQ2_M | 11.08 GB | 2.89 |
gemma-4-31B-it-DASHQ-Q2_K_XL.gguf |
Q2_K_XL | 11.98 GB | 3.12 |
Text only: the vision tower is not included.
Quality vs. F16 (KL divergence)
Raw-text perplexity is not a meaningful metric for gemma-4-it (the F16 model itself scores in the thousands on WikiText-2 while generating normally),
so quality is reported as KL divergence of the next-token distribution to F16 (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) | 8.66 GB | 5.49 | 4.36 | 25.8% |
| IQ2_XXS | unsloth UD-IQ2_XXS | 8.53 GB | 4.92 | 3.73 | 29.4% |
| IQ2_XXS | DASH-Q IQ2_XXS | 8.82 GB | 3.66 | 2.02 | 41.5% |
| IQ2_XS | llama.cpp IQ2_XS (imatrix) | 9.51 GB | 4.88 | 3.64 | 29.5% |
| IQ2_XS | DASH-Q IQ2_XS | 10.03 GB | 2.53 | 1.18 | 51.9% |
| IQ2_M | llama.cpp IQ2_M (imatrix) | 10.90 GB | 3.38 | 2.25 | 39.4% |
| IQ2_M | unsloth UD-IQ2_M | 10.75 GB | 2.18 | 0.90 | 55.5% |
| IQ2_M | DASH-Q IQ2_M | 11.08 GB | 1.41 | 0.37 | 63.8% |
| Q2_K_XL | llama.cpp Q2_K (imatrix) | 11.92 GB | 3.04 | 1.96 | 44.4% |
| Q2_K_XL | DASH-Q (this file) | 11.98 GB | 1.38 | 0.32 | 65.8% |
Usage
llama-cli -m gemma-4-31B-it-DASHQ-IQ2_M.gguf -ngl 99 -c 8192
License
Inherits the license of the base model (google/gemma-4-31B-it).