Qwen3.5-27B - DASH-Q 2-bit GGUF
llama.cpp GGUF files of Qwen/Qwen3.5-27B 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 |
|---|---|---|---|
Qwen3.5-27B-DASHQ-IQ2_XXS.gguf |
IQ2_XXS | 8.58 GB | 2.51 |
Qwen3.5-27B-DASHQ-IQ2_XS.gguf |
IQ2_XS | 9.37 GB | 2.74 |
Qwen3.5-27B-DASHQ-IQ2_M.gguf |
IQ2_M | 10.07 GB | 2.94 |
Qwen3.5-27B-DASHQ-Q2_K_XL.gguf |
Q2_K_XL | 10.81 GB | 3.16 |
Text only: the vision tower is not included.
Perplexity (lower is better)
| Type | Model | Size | WikiText-2 | C4 |
|---|---|---|---|---|
| Q8_0 | 29.0 GB | 6.62 | 9.63 | |
| IQ2_XXS | llama.cpp IQ2_XXS (imatrix) | 8.67 GB | 10.67 | 11.72 |
| IQ2_XXS | unsloth UD-IQ2_XXS | 8.57 GB | 7.94 | 11.35 |
| IQ2_XXS | DASH-Q IQ2_XXS | 8.58 GB | 8.24 | 11.19 |
| IQ2_XS | llama.cpp IQ2_XS (imatrix) | 9.33 GB | 8.49 | 11.06 |
| IQ2_XS | DASH-Q IQ2_XS | 9.37 GB | 7.49 | 10.53 |
| IQ2_M | llama.cpp IQ2_M (imatrix) | 10.24 GB | 8.06 | 10.56 |
| IQ2_M | unsloth UD-IQ2_M | 10.19 GB | 7.13 | 10.38 |
| IQ2_M | DASH-Q IQ2_M | 10.07 GB | 7.12 | 10.22 |
| Q2_K_XL | llama.cpp Q2_K (imatrix) | 10.86 GB | 7.09 | 10.49 |
| Q2_K_XL | unsloth UD-Q2_K_XL | 11.21 GB | 7.45 | 10.24 |
| Q2_K_XL | DASH-Q Q2_K_XL | 10.81 GB | 7.12 | 10.15 |
llama-perplexity, context 2048; WikiText-2 test, C4 validation (256 x 2048 tokens).
Zero-shot accuracy (lm-eval-harness, 0-shot)
| File | arc_easy | arc_challenge | piqa | hellaswag | winogrande | openbookqa | commonsense_qa | truthfulqa_mc2 | lambada_openai | avg |
|---|---|---|---|---|---|---|---|---|---|---|
| DASH-Q Q2_K_XL | 80.4 | 62.2 | 81.6 | 80.9 | 78.6 | 43.0 | 85.4 | 56.1 | 76.0 | 71.58 |
Usage
llama-cli -m Qwen3.5-27B-DASHQ-Q2_K_XL.gguf -ngl 99 -c 8192
License
Inherits the license of the base model (Qwen/Qwen3.5-27B).