dQwen3.5-9B-Base

A masked diffusion language model adapted from Qwen3.5-9B. The backbone is hybrid: only its attention layers are made bidirectional, and the Gated DeltaNet layers stay causal. This is a base model, with no instruction tuning.

Paper: dQwen3.5: Hybrid-Attention Diffusion Language Models
Code: https://github.com/AntonXue/dQwen

The dQwen3.5 family

Quickstart

Needs a CUDA GPU and transformers>=5.13 (tested with torch 2.7.1+cu128, flash-linear-attention 0.5.1).

import torch
from transformers import AutoModel

model = AutoModel.from_pretrained("UT-IFML/dQwen3.5-9B-Base", trust_remote_code=True, dtype=torch.bfloat16).cuda().eval()
stop_strings = ["\n\n", "\ndef", "\nclass", "<|endoftext|>"]
print(model.generate("def fibonacci(n):\n", gen_length=512, stop_strings=stop_strings).text)

generate decodes the whole canvas at once, committing positions above a confidence threshold (conf_threshold=0.9, the paper's tau); pass block_length=32 for left-to-right block decoding, or conf_threshold=None, steps_per_block=k for a fixed budget. A position's confidence is its top-1 probability (conf_source="top1"); conf_source="sampled" reads the sampled token's probability instead, the default before 2026-11-11. The two differ only when sampling (temperature>0): greedy output, and every result in the paper, is unchanged. tau= and order= still work as deprecated names for conf_threshold= and remask_order=. The 50B-token checkpoint from the paper is revision="step25000-swa".

Citation

@article{xue2026dqwen,
  title   = {dQwen3.5: Hybrid-Attention Diffusion Language Models},
  author  = {Xue, Anton and Rout, Litu and Akella, Aditya and Klivans, Adam and Sanghavi, Sujay and Shakkottai, Sanjay},
  journal = {arXiv preprint arXiv:2609.20751},
  year    = {2026}
}