FlashMap
A speculative decoding drafter based on a flow map that jointly drafts 16 tokens in a single forward pass, substantially increasing LLM decoding speed while keeping the output distribution lossless.
Timeline (last 90 days)
Polymathic AI, together with researchers from Princeton University and USC ISI, released FlashMap, a speculative decoding drafter based on flow mapping.
FlashMap achieves up to 6.14x lossless speedup on Qwen3-8B.
FlashMap can jointly draft 16 tokens in a single forward pass.
The 6.14x speedup claimed by FlashMap is achieved under a lossless constraint, i.e., the output distribution is identical to that of the target model decoding alone.
6.14x is the highest speedup measured on the specific model Qwen3-8B; the report does not provide complete curves across models or tasks, so it should be understood as a peak value under a specific configuration rather than a universal end-to-end throughput improvement.
At present, there is no clear word yet on whether FlashMap will be open-sourced, its adaptation cost, or its integration path with existing inference frameworks (such as vLLM and SGLang).
FlashMap shares the same goal as Medusa-style multi-head prediction and EAGLE-style feature-level drafting; the difference lies in the mechanism by which drafts are generated.
All Facts (7)
The 6.14x speedup claimed by FlashMap is achieved under a lossless constraint, i.e., the output distribution is identical to that of the target model decoding alone.
50%UnverifiedFlashMap achieves up to 6.14x lossless speedup on Qwen3-8B.
50%UnverifiedFlashMap can jointly draft 16 tokens in a single forward pass.
50%UnverifiedFlashMap shares the same goal as Medusa-style multi-head prediction and EAGLE-style feature-level drafting; the difference lies in the mechanism by which drafts are generated.
50%Unverified6.14x is the highest speedup measured on the specific model Qwen3-8B; the report does not provide complete curves across models or tasks, so it should be understood as a peak value under a specific configuration rather than a universal end-to-end throughput improvement.
50%UnverifiedAt present, there is no clear word yet on whether FlashMap will be open-sourced, its adaptation cost, or its integration path with existing inference frameworks (such as vLLM and SGLang).
50%UnverifiedPolymathic AI, together with researchers from Princeton University and USC ISI, released FlashMap, a speculative decoding drafter based on flow mapping.
50%