DSpark

DeepSeek open-sourced DSpark, a speculative decoding framework that bolts a draft module onto existing DeepSeek-V4 weights. No retraining the base model. The draft module does the heavy lifting, proposing tokens that the main model then verifies.

The architecture pairs a parallel draft backbone with a lightweight Markov head. That combination cuts the suffix decay problem that plagues simpler speculative approaches. On top of that, confidence-scheduled verification adjusts how many tokens get checked based on live GPU load, which matters a lot in real production environments.

The numbers are hard to ignore. Offline, accepted token length rises 16 to 31% over DFlash and Eagle3. In production, per-user generation speed improves 57 to 85% over the MTP-1 baseline, with no quality loss according to DeepSeek’s published results.

Speculative decoding has been a known technique for a while, but making it work cleanly in high-traffic serving setups is the hard part. DSpark looks like a serious attempt at that, and the training repo is public.