deepseek-ai/deepspec

Large language models are powerful, but generating text token by token can be slow. A technique called "speculative decoding" speeds this up by using a smaller, faster "draft" model to guess what the larger "target" model will say next. If the guesses are right, the large model can accept multiple words at once instead of computing them one by one. DeepSpec is a toolkit for training and evaluating those draft models, so teams can make their AI inference faster without changing the final output quality. The project provides a three-stage pipeline. First, you prepare training data by running a target model to generate answers and caching its internal states. Second, you train a draft model to predict those cached outputs. Third, you evaluate how well the draft model's guesses are accepted across standard benchmarks like math, coding, and general conversation tasks. The training script assumes a machine with 8 GPUs, and the data preparation step can require roughly 38 terabytes of storage for the default configuration. This is aimed at AI researchers and engineers who want to experiment with or compare speculative decoding algorithms. For example, someone deploying a Qwen3 or Gemma model in production could use this to train a companion draft model that accelerates response times. The repo includes pre-trained checkpoints for three algorithms, DSpark, DFlash, and Eagle3, paired with four target models, so users can also evaluate existing approaches without training from scratch. A few things stand out. The resource requirements are substantial: the default setup expects multiple GPUs and massive storage for the target cache, putting this firmly in the realm of well-resourced teams rather than hobbyists. The project also builds on existing open-source work, the training framework comes from SpecForge, the DFlash design from its own repo, and the authors encourage others to contribute new algorithms. For domain-specific applications, they note that fine-tuning the draft model on your own data is recommended, especially if the target model uses a "thinking" mode.