Researchers from Polymathic AI, Princeton and NYU have released FlashMap, a speculative decoding drafter that swaps a discrete diffusion model for a continuous flow map. It drafts 16 tokens in a single forward pass, with a reported lossless speedup of up to 6.14x on GSM8K with Qwen3-8B. Weights for Qwen3-8B and 4B are on Hugging Face, with vLLM and SGLang plugins. https://www.marktechpost.com/2026/10/28/flashmap-flow-map-speculative-decoding/ #AIagent #AI #GenAI #AIresearch
MarkTechPostPolymathic AI and Princeton Researchers Release FlashMap, a Flow Map Drafter for 6.14x Faster LLM DecodingFlashMap speculative decoding uses a 0.5B flow map drafter to reach 6.14x lossless speedup on Qwen3-8B, beating DFlash.