| library_name: exllamav3 |
| tags: |
| - exl3 |
| - exllamav3 |
| - quantized |
| - dflash |
| - speculative-decoding |
| --- |
| |
| # Gemma 4 12B IT DFlash (EXL3) |
| |
| This is an **ExLlamaV3 (EXL3)** quantized version of the `z-lab/gemma4-12B-it-DFlash` model. |
| |
| DFlash models are specifically designed to act as **Draft Models** for Speculative Decoding, allowing you to generate text significantly faster when paired with a larger main model. These were quantized using the Ampere-optimized `mul1` codebook for maximum decoding speed on modern GPUs. |
| |
| ## 📥 Available Branches / Downloads |
| |
| The model weights are stored in separate branches. **Please click a branch link below to view and download the specific size you need.** |
| |
| | Target BPW | Head BPW | Branch (Download Link) | |
| |:---|:---:|:---| |
| | **6.00** | 6 | [6.0bpw_h6](https://huggingface.co/DBMe/gemma4-12B-it-DFlash-exl3/tree/6.0bpw_h6) | |
| | **5.00** | 6 | [5.0bpw_h6](https://huggingface.co/DBMe/gemma4-12B-it-DFlash-exl3/tree/5.0bpw_h6) | |
| | **4.00** | 6 | [4.0bpw_h6](https://huggingface.co/DBMe/gemma4-12B-it-DFlash-exl3/tree/4.0bpw_h6) | |
| | **3.50** | 6 | [3.5bpw_h6](https://huggingface.co/DBMe/gemma4-12B-it-DFlash-exl3/tree/3.5bpw_h6) | |
| | **3.00** | 6 | [3.0bpw_h6](https://huggingface.co/DBMe/gemma4-12B-it-DFlash-exl3/tree/3.0bpw_h6) | |
| | **2.50** | 6 | [2.5bpw_h6](https://huggingface.co/DBMe/gemma4-12B-it-DFlash-exl3/tree/2.5bpw_h6) | |
| |
| ## 💻 How to Download |
| |
| It is highly recommended to use the `huggingface-cli` to download a specific branch to avoid downloading the entire repository (which contains all sizes). |
| |
| ```bash |
| pip install -U "huggingface_hub[cli]" |
| ``` |
| |
| **Example: Downloading the 4.0 BPW branch** |
| ```bash |
| huggingface-cli download DBMe/gemma4-12B-it-DFlash-exl3 \ |
| --revision 4.0bpw_h6 \ |
| --local-dir gemma4-12B-it-DFlash-exl3-4.0bpw |
| ``` |
| |
| ## ⚙️ Quantization Details |
| * **Engine:** ExLlamaV3 |
| * **Codebook:** `mul1` (Ampere-optimized) |
| * **Head BPW:** 6.0 (Preserves output layer accuracy for drafting) |
| * **Calibration:** Uncalibrated (Standard practice for DFlash architectures) |
