DSpark: Speculative decoding accelerates LLM inference [pdf]
Points and comments are a snapshot, not live.
DeepSeek's DSpark speculative decoding speeds LLM inference in production.
DSpark uses speculative decoding, where a smaller draft model generates multiple tokens that a larger target model then verifies in parallel. DeepSeek co-deployed DSpark draft models with DeepSeek-V4-Flash and DeepSeek-V4-Pro. Section 5.4 notes MTP-1 (the prior production setup) was superseded by DSpark two weeks after the DeepSeek-V4-preview release. At matched system capacities, DSpark delivers 57% to 78% faster per-user generation.
The paper is hosted on GitHub under the DeepSpec repository by DeepSeek AI.
What commenters are saying
Commenters widely praise DeepSeek's openness and practical efficiency, contrasting it with US labs' secrecy and capital-intensive approaches. Several note DSpark likely contributed to DeepSeek's recent price cuts. A sub-thread debates whether Chinese labs publish out of necessity (marketing and catching up) or cultural tradition; some predict they will stop if they take the lead. Others argue US labs' secrecy hurts their own progress, citing historical cryptography examples where public research outpaced closed efforts. One commenter reports using DeepSeek V4 Pro in Kilo Code for 1.5B tokens at $40.