DeepSpec is a full-stack codebase released by deepseek-ai for training and evaluating draft models used in speculative decoding. The project provides data preparation utilities, implementation code, and evaluation scripts to facilitate the development of these auxiliary models.
The collection includes checkpoints for three supported algorithms: DSpark, DFlash, and Eagle3. Pre-trained weights are available for target models including Qwen/Qwen3-4B, Qwen/Qwen3-8B, Qwen/Qwen3-14B, and google/gemma-4-12B-it.
Users are advised to align their setup with the repository's training settings to ensure meaningful comparisons. For domain-specific applications, fine-tuning the draft model is recommended, particularly if the target model operates in thinking mode.