FlashMap: Polymathic AI and Princeton release a speculative decoding drafter for LLMs
Overview
Researchers from Polymathic AI, Princeton University, USC ISI, NYU and the Flatiron Institute release FlashMap, a speculative decoding drafter that proposes a 16-token block in one forward pass.
The research team reports up to a 6.14x lossless speedup on GSM8K with Qwen3-8B.
The article compares FlashMap with DFlash, saying the drafter has about half the parameters and beats it on acceptance length in 15 of 18 one-step settings. Drafter weights for Qwen3-8B and Qwen3-4B are on Hugging Face, with vLLM and SGLang plugins.
Written by AI from the articles below · updated Oct 28, 4:54 AM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- MarkTechPostNewsPickPolymathic AI and Princeton researchers release FlashMap, a flow map drafter for faster LLM decoding
AIResearchers from Polymathic AI, Princeton University, and other institutions released FlashMap, a speculative decoding drafter that proposes a 16-token block in one forward pass. The research team reports up to a 6.14x lossless speedup on GSM8K with Qwen3-8B, with drafter weights on Hugging Face and vLLM and SGLang plugins. The article's comparison notes that the drafter has about half the parameters of DFlash and beats it on acceptance length in 15 of 18 one-step settings.
Heat Trend
Not enough continuous observations to show a trend yet.