Researchers from Princeton University, Polymathic AI, USC ISI, NYU, and the Flatiron Institute have released FlashMap, a new speculative decoding drafter that delivers up to a 6.14x lossless speedup on Qwen3-8B. FlashMap replaces discrete diffusion drafters with a continuous flow map, enabling the generation of a full 16-token block in a single forward pass while maintaining token consistency. The research team tested the system across nine benchmarks covering math, code, and chat at temperatures zero and one, finding that the one-step FlashMap outperforms retrained baselines like DFlash in acceptance length and overall execution speed.
The architecture utilizes a 0.5B-parameter drafter—specifically 500M parameters for Qwen3-8B and 475M for Qwen3-4B—running on the same GPU as the target model. Built upon Flow Map Language Models, FlashMap maps Gaussian noise directly to a token block in one step, allowing the noise to jointly pick the mode for every slot and avoiding the independent slot prediction errors seen in traditional block diffusion drafters. Training involved a two-stage process consisting of a flow teacher trained with cross-entropy and a flow map student using semigroup distillation, trained on 500K samples from Open-PerfectBlend.
The open-source release includes drafter weights for Qwen3-8B and Qwen3-4B available on Hugging Face, alongside vLLM and SGLang plugins for deployment. The flashmap package requires PyTorch and Transformers, enabling straightforward integration for developers seeking to accelerate large language model inference without sacrificing output accuracy. This advancement significantly reduces token generation latency, improving the practical efficiency of large-scale AI deployments.