FlashMap can speed up LLM replies up to 6.14x on Qwen3-8B
FlashMap is a new drafter from Polymathic AI, Princeton and USC ISI researchers that guesses 16 tokens at once to speed up language model replies. It reaches up to 6.14x lossless speedup on Qwen3-8B. It is aimed at AI developers running models, not everyday users.
How to try it
This is a developer tool, so start with the release page from its researchers. The story does not say where to download it.
Useful 15%