On 2026-06-26 DeepSeek published github.com/deepseek-ai/DeepSpec, a full-stack MIT-licensed codebase for training and evaluating draft models for speculative decoding. The bundle ships three implementations — DSpark (new, with a paper in the repo), DFlash (block diffusion, accepted at ICML 2026, with the DeepSeek authors claiming over 6× lossless acceleration and up to 2.5× over EAGLE-3), and Eagle3 (the prior state of the art) — with 8-GPU training scripts, 38 TB of target-cache storage for the Qwen3-4B default, and evaluation across nine benchmarks. The 1,546 stars, 128 forks, single commit on main, and zero release tags as of 2026-06-28 are the load-bearing repo signals. Every speedup number in the article is the authors' own measurement — no independent reproduction is on the record as of 2026-06-28.
DeepSeek releases DeepSpec: open-source full-stack for training and evaluating speculative decoding draft models