| GUI Planner DFlash2 Draft | |
| This release contains fully finetuned draft weights derived from: | |
| z-lab/Qwen3.8-27B-DFlash2 | |
| Revision: 50307d4c4cde6860d4eee73e2547cd786fe8e8a4 | |
| https://huggingface.co/z-lab/Qwen3.8-27B-DFlash2/tree/50307d4c4cde6860d4eee73e2547cd786fe8e8a4 | |
| The upstream draft is distributed under Apache License 2.0. Original work | |
| and implementation are attributed to the upstream model and DFlash project. | |
| Changes in this release: all draft backbone, dynamic-convolution and | |
| candidate-selector parameters were finetuned for GUI planning for two | |
| epochs, then exported as BF16 safetensors. The original block-8 configuration | |
| is preserved; block 12 is selected only through inference configuration. | |
| The compatible target, tokenizer, processor and datasets are separate assets. | |
| The optional serve_planner.py helper applies the same Qwen XML whitespace | |
| policy used in the reported benchmark. It is provided under Apache 2.0. | |