GLM-4.7-Flash - DASH-Q 2-bit GGUF
llama.cpp GGUF files of zai-org/GLM-4.7-Flash 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 |
|---|---|---|---|
GLM-4.7-Flash-DASHQ-IQ2_XXS.gguf |
IQ2_XXS | 10.25 GB | 2.63 |
GLM-4.7-Flash-DASHQ-IQ2_M.gguf |
IQ2_M | 10.73 GB | 2.76 |
GLM-4.7-Flash-DASHQ-Q2_K_XL.gguf |
Q2_K_XL | 11.55 GB | 2.97 |
Perplexity (lower is better)
| Type | Model | Size | WikiText-2 | C4 |
|---|---|---|---|---|
| Q8_0 | 33.20 GB | 10.37 | 17.51 | |
| IQ2_XXS | llama.cpp IQ2_XXS (imatrix) | 8.77 GB | 31.83 | 38.15 |
| IQ2_XXS | unsloth UD-IQ2_XXS | 10.51 GB | 9.51 | 20.37 |
| IQ2_XXS | DASH-Q IQ2_XXS | 10.25 GB | 12.58 | 15.99 |
| IQ2_XS | llama.cpp IQ2_XS (imatrix) | 9.66 GB | 14.63 | 21.89 |
| IQ2_M | llama.cpp IQ2_M (imatrix) | 10.69 GB | 10.13 | 15.86 |
| IQ2_M | unsloth UD-IQ2_M | 10.99 GB | 12.04 | 20.27 |
| IQ2_M | DASH-Q IQ2_M | 10.73 GB | 8.88 | 13.92 |
| Q2_K_XL | llama.cpp Q2_K (imatrix) | 11.82 GB | 16.46 | 16.39 |
| Q2_K_XL | unsloth UD-Q2_K_XL | 11.89 GB | 12.55 | 15.93 |
| Q2_K_XL | DASH-Q Q2_K_XL | 11.56 GB | 12.17 | 13.99 |
llama-perplexity, context 2048; WikiText-2 test, C4 validation (256 x 2048 tokens).
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) | 8.69 GB | 4.44 | 3.55 | 23.1% |
| IQ2_XXS | unsloth UD-IQ2_XXS | 10.51 GB | 0.22 | 0.09 | 81.8% |
| IQ2_XXS | DASH-Q IQ2_XXS | 10.25 GB | 0.22 | 0.09 | 81.4% |
| IQ2_XS | llama.cpp IQ2_XS (imatrix) | 9.59 GB | 1.82 | 1.28 | 45.8% |
| IQ2_M | llama.cpp IQ2_M (imatrix) | 10.63 GB | 0.56 | 0.32 | 69.4% |
| IQ2_M | unsloth UD-IQ2_M | 10.99 GB | 0.21 | 0.08 | 82.5% |
| IQ2_M | DASH-Q IQ2_M | 10.73 GB | 0.20 | 0.07 | 82.4% |
| Q2_K_XL | llama.cpp Q2_K (imatrix) | 11.75 GB | 0.69 | 0.40 | 66.5% |
| Q2_K_XL | unsloth UD-Q2_K_XL | 11.89 GB | 0.19 | 0.07 | 83.1% |
| Q2_K_XL | DASH-Q Q2_K_XL | 11.55 GB | 0.18 | 0.06 | 84.1% |
Usage
llama-cli -m GLM-4.7-Flash-DASHQ-Q2_K_XL.gguf -ngl 99 -c 8192
License
Inherits the license of the base model (zai-org/GLM-4.7-Flash).