kun432🇯🇵

Diffusion-Qwen3.8-27B is officially out. I adapted Qwen3.8-27B into a masked diffusion language model. The prompt stays fixed, and generation starts with a sequence of masked tokens. On each pass, the model predicts multiple missing tokens in parallel and commits those that meet a confidence threshold. The next pass uses those tokens as context to resolve more of the response. To support this, I removed the causal mask from the full-attention layers so each position can use context on both sides. The Gated DeltaNet layers scan the sequence in both directions, with a learned gate controlling how much the reverse scan contributes. This is a research checkpoint. I haven't benchmarked its accuracy or generation speed against the base model yet. One possible failure mode is an early, confident mistake. If the model commits to the wrong word, later passes can build a fluent but incorrect answer around it. The weights, code, and a generation example are public. If you hit a failure, please share the prompt, decoding settings, and raw output so I can reproduce it. huggingface.co/summerAIO/Dif…