Researchers from Polymathic AI, Princeton University, USC ISI, NYU and the Flatiron Institute have released FlashMap, a “drafter” designed for speculative decoding of language models. According to the team, it can accelerate generation on Qwen3-8B by up to 6,14 times without altering the output. The drafter weights for Qwen3-8B and Qwen3-4B are available on Hugging Face.
- FlashMap accelerates language model generation by up to 6.14 times.
- It drafts a block of 16 tokens at once, keeping them consistent.
- It outperforms DFlash on acceptance length with about half its parameters.
- Its weights are released on Hugging Face under Apache-2.0.
Speculative decoding involves having a small auxiliary model propose multiple tokens (the text units processed by a model) which the large model then verifies in a single pass. FlashMap proposes a full block of 16 tokens at once, and according to the authors, the tokens in this block remain consistent with each other. The drafter counts 0,5 milliard parameters, i.e., 500 millions for Qwen3-8B and 475 millions for Qwen3-4B, and runs on the same GPU as the target model.
Why do diffusion drafters lose precision?
The DFlash drafter predicts each position in the block independently, which the researchers call a factorized approximation. The source gives the example of a word sequence that could be “new york” or “san diego” with equal probabilities. An independent drafter then proposes “new diego” or “san york” half the time, and the target model rejects these sequences at the second position.
FlashMap builds on Flow Map Language Models. Each noise sample corresponds to a full block, and this noise jointly determines the selected version for each position. The drafter is a diffusion transformer (DiT) with 6 blocks, an embedding dimension of 2048 and 16 attention heads. It reads the hidden states of layers 1,9, 17,25 and 33 of the target model.
Results announced on nine benchmarks
The researchers retrained EAGLE-3 and DFlash on the same data and under the same conditions, then evaluated the drafters on nine benchmarks covering mathematics, code and chat, at temperatures 0 and 1. According to the source, FlashMap in a single step is the fastest drafter on every task tested on Qwen3-8B, at both temperatures. It outperforms DFlash on acceptance length in 15 out of 18 configurations, with roughly half its parameters.
In two-step mode, acceptance length surpasses DFlash in 18 out of 18 configurations, with an average improvement of 10 %. The authors also note that on Qwen3-4B at temperature 0, FlashMap is beaten by DFlash on 4 of the 9 tasks. The lowest reported case is a 2,28‑fold gain on Alpaca at temperature 1. The model was trained on 500 000 samples from Open‑PerfectBlend, for a total of roughly 4 600 compute hours on H200 GPUs.
For deployment, the flashmap package requires only torch and transformers version 4.56 or higher, and the projects provide usage paths via vLLM 0.31 and SGLang 0.5.20. The weights are released under the Apache‑2.0 license and the code under the MIT license.
