FlashMap drafter delivers up to 6.14x faster lossless LLM decoding

Polymathic AI, Princeton University and partners released FlashMap, a speculative decoding drafter that uses a continuous flow map to propose a 16-token block in one forward pass. On Qwen3-8B with a 0.5B-parameter drafter it reports up to 6.14x lossless speedup on GSM8K. Weights for Qwen3-8B and Qwen3-4B are on Hugging Face with vLLM and SGLang plugins.

• 0.5B-parameter drafter proposes 16 tokens in a single pass
• Up to 6.14x speedup on GSM8K at temperature 0
• Faster than retrained DFlash on all 9 Qwen3-8B benchmarks
• Apache-2.0 weights, MIT code, plugins for vLLM 0.31 and SGLang 0.5.20