al-jshen/flashmap is a drafter for speculative decoding. Speculative decoding is a method where a small model proposes the next tokens first and a large model verifies those proposals, so that several tokens are confirmed at once. FlashMap takes on that stage as a drafter that operates without loss. It is written in Python and follows the MIT license. There are no public releases, and the repository, created on October 13, 2026, released the paper, Hugging Face weights, and SGLang and vLLM integrations together. It has 19 GitHub stars.
The difference comes down to whether tokens are predicted one at a time. Diffusion-style drafters predict each token independently without using earlier results, so they miss the dependencies between tokens. FlashMap uses a flow map approach, that is, it generates the entire block at once from noise, and captures dependencies inside the block as well. The README also contains an implementation for Transformers, a tool for converting trained checkpoints, and a drafter for Qwen models together with a description of its training procedure.
That said, the performance evidence stays at the level of a README claim. The README states that it delivers up to 6x speedup, but that figure cannot be verified independently, and no dedicated test directory can be found in the repository. The scores are novelty 9, tech 6, quality 6, and momentum 2.2, out of 10. Taken together with 19 stars and the low momentum, this is still an unverified early proposal. At the moment it is closer to the stage where one would directly check the SGLang or vLLM integration than to trusting the README numbers.