FlashMap Achieves 6.14x Faster LLM Decoding via Flow Map Speculative Decoding

Researchers from Polymathic AI and Princeton have released FlashMap, a speculative decoding framework that uses flow maps to draft token sequences and achieves up to 6.14x faster LLM decoding without degrading output quality. Speculative decoding remains one of the most practical inference acceleration techniques, and FlashMap's flow map approach offers a new architectural angle on the draft-then-verify paradigm. The key innovation is using learned flow maps to generate high-quality draft tokens that are more likely to be accepted by the verifier model, reducing wasted compute. Developers running inference at scale — particularly those constrained by latency budgets — should evaluate FlashMap as a drop-in acceleration layer. The research is publicly available, making it actionable for teams building custom inference stacks.