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

Polymathic AI and collaborators released FlashMap, a speculative decoding drafter that speeds up large language model output by up to 6.14x on Qwen3-8B. FlashMap drafts 16-token blocks in one pass, improving consistency and efficiency over previous methods.