DeepSeek launches DSpark framework to boost AI inference speed by up to 85%

Chinese AI start‑up DeepSeek announced the DSpark speculative decoding framework as part of its V4 model upgrade. DSpark combines a lightweight draft model with batch verification by a larger model, using semi‑autoregressive generation and a confidence‑scheduled scheduler. The technique speeds single‑user response generation by 60‑85% and can double the number of queries a GPU handles (from about 100 to roughly 185). In live traffic DSpark reduced latency and cut computational costs, delivering a 51% increase in overall throughput in stringent latency scenarios.

DeepSeek released the accompanying DeepSpec training toolkit and model weights under an MIT licence, hosting code on GitHub and HuggingFace. The framework was tested on open‑source models such as Qwen3, Gemma and others, and has been deployed in the V4‑Flash and V4‑Pro preview services. The upgrade also introduces a peak‑hour pricing model that doubles API rates during high‑traffic periods. The release comes as Chinese developers seek to lower inference costs amid tightening U.S. chip export restrictions.

A joint paper with Peking University details the semi‑autoregressive design and confidence‑based token verification, emphasizing that the acceleration incurs no loss of output quality.