ChainFlow preprint reports faster speculative decoding on Qwen3.5 models

Researchers introduced ChainFlow, a method for accelerating language-model inference, and reported that it outperforms EAGLE-3 by 39% and DFlash by 11% in vLLM tests on Qwen3.5 models. The preprint compares systems with matched drafter size and training data. Speculative decoding uses a smaller model to propose tokens that the target model checks together. ChainFlow drafts through one or two transformations of compressed internal model states, with feedback components that preserve dependencies between tokens and tree drafting that raises acceptance. The researchers say they are open-sourcing code and drafter models. The reported comparisons cover Qwen3.5 models under the study's test conditions.