Polymathic AI and Princeton Researchers Release FlashMap, a Flow Map Drafter for 6.14x Faster LLM Decoding

A team from Polymathic AI, Princeton, USC ISI, NYU and the Flatiron Institute has released FlashMap, a speculative decoding drafter that swaps block diffusion for a continuous flow map. Instead of guessing tokens one by one, it drafts a full 16-token block in one forward pass and lets the target model verify the whole block in parallel.

The hook is simple: keep the speed, lose the usual quality hit. The paper reports that FlashMap can reach up to 6.14x lossless speedup on GSM8K with Qwen3-8B, using a 0.5B-parameter drafter. The authors also say it is the fastest drafter they tested on Qwen3-8B, coming in about 7% ahead of a retrained DFlash on average.

Why does that matter? Because diffusion-style drafters often hurt acceptance when they treat each slot in a block as independent. FlashMap is built around Flow Map Language Models, where one noise sample maps to one whole block, so the model can choose the mode for all tokens together instead of producing awkward mixtures like “new diego” or “san york.” The team says that joint structure is what helps acceptance stay higher.

The drafter itself is a 6-block diffusion transformer with 2048-dimensional embeddings and 16 attention heads. It pulls hidden states from target layers 1, 9, 17, 25 and 33, while the embedding matrix and LM head are copied from the target and frozen. Training happens in two stages: a flow teacher first, then a student trained by semigroup distillation to jump from noise to data in one step.

On benchmarks, the team retrained EAGLE-3 and DFlash on the same data and settings. Across nine tasks spanning math, code and chat, FlashMap was the fastest drafter on every Qwen3-8B test at both temperatures. It beat DFlash on acceptance length in 15 of 18 settings, and a two-step mode improved that to 18 of 18. The paper also says the weights for Qwen3-8B and Qwen3-4B are on Hugging Face, with plugins for vLLM and SGLang.