DeepSeek Open-Sources DSpark and DeepSpec to Speed Model Inference

Overview

• DeepSeek published the DSpark speculative-decoding framework, the DeepSpec training and evaluation toolchain, and two DSpark-augmented DeepSeek-V4 model builds on GitHub and Hugging Face.
• The company says DSpark uses a semi-autoregressive draft generator plus a confidence-scheduled verification step to reduce wasted verification work and preserve token dependencies during draft generation.
• DeepSeek reports single-user speed gains of 60%–85% versus its production baseline when DSpark runs on DeepSeek-V4 but the coverage notes these performance claims come from DeepSeek and the paper without independent third-party validation.
• DeepSpec is released under an MIT license and includes data preparation, draft-model implementations, training code, and evaluation scripts so other teams can train and test speculative-decoding drafts for existing models.
• The open-source release is framed as a strategic move to lower inference cost and speed product rollout after a reported large financing and includes acknowledgements to projects such as SpecForge, Qwen3, and Gemma.