DeepSeek Open-Sources DSpark: Semi-Autoregressive Speculative Decoding Boosts Throughput 51–400%

DeepSeek (GitHub) / Hacker News·medium signalXBlueskyLinkedInCopy link

DeepSeek released DSpark ('Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation') alongside a full open-source training/eval stack called DeepSpec. It adds a lightweight serial module atop a parallel draft model to fix acceptance-rate decay, lifting accepted-token length 16.3–30.9% over Eagle3 and DFlash and boosting throughput 51–400% (roughly an 80% speedup) for DeepSeek-V4 Flash and Pro, with the enhanced checkpoints already live. The paper PDF was the #1 story on Hacker News at 510 points.