DSpark / DeepSpec

DSpark ("Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation") is DeepSeek's production speculative-decoding framework: a semi-autoregressive draft architecture (parallel backbone plus a lightweight sequential module that mitigates the "suffix decay" of parallel drafters) combined with confidence-scheduled verification that dynamically truncates draft length per request. In offline evaluation it beats EAGLE-3 by ~27–31% and DFlash (arXiv 2602.06036) by ~16–18% macro-average accepted length on Qwen3-4B/8B/14B; deployed on live DeepSeek-V4 traffic it raises per-user generation speed 57–85% over the prior MTP-1 baseline at matched aggregate throughput. The paper ships as a PDF in the repo (not on arXiv), following the V3-era pattern.

DeepSpec is the accompanying MIT-licensed full-stack codebase for training and evaluating draft models (data prep, training, evaluation), implementing DSpark alongside DFlash and EAGLE-3 baselines. DeepSeek also re-uploaded DeepSeek-V4 Pro and Flash as "-DSpark" checkpoints — identical weights with the draft module attached — plus a 12-checkpoint draft-model zoo covering Qwen3 and Gemma 4 targets. Ecosystem adoption was immediate: Red Hat trained a GLM-5.2 DSpark speculator within a week of release.