Model releases & benchmarks
dQwen3.5: Hybrid-Attention Diffusion Language Models
Can an autoregressive hybrid with causal RNN layers be adapted into a diffusion language model? The authors report that it can, with a matched-loss token result.
Why it matters: The reported result suggests hybrid pretrained backbones may be efficient starting points for DLM adaptation when training-token use matters. The abstract does not establish lower wall-clock training time, faster serving, or a universal quality advantage.