# DeepSeek and Peking University Open-Source Speculative Decoding Acceleration Framework DSpark

- Category: Flash (快讯)
- Published: 2026-06-29T11:58:00.000Z
- Source date: 2026-06-29
- HTML: https://www.88z.ai/en/news/flash-20260629-46942952d33564c3
- Markdown: https://www.88z.ai/en/news/flash-20260629-46942952d33564c3.md
- Other language: https://www.88z.ai/news/flash-20260629-46942952d33564c3.md

DeepSeek and Peking University have jointly released DSpark, a large model inference acceleration framework now deployed in the DeepSeek-V4 production system. While maintaining total throughput, DSpark significantly boosts single-user generation speed: DeepSeek-V4-Flash sees a 60%–85% increase, and DeepSeek-V4-Pro a 57%–78% rise. It utilizes a semi-autoregressive architecture with confidence-based scheduling verification to balance draft generation efficiency and text coherence, dynamically adju

## At a glance

- DSpark is deployed in the DeepSeek-V4 production system, significantly boosting single-user generation speed.
- It uses a semi-autoregressive architecture with confidence scheduling, supporting dynamic verification length adjustment based on load.

## The update

DeepSeek, in collaboration with Peking University, has launched DSpark, a large model inference acceleration framework now integrated into the DeepSeek-V4 production environment. Data shows that while maintaining stable total throughput, DSpark effectively optimizes the single-user response experience. Specifically, the single-user generation speed for the DeepSeek-V4-Flash model has increased by 60% to 85%, while the DeepSeek-V4-Pro model has seen an improvement of 57% to 78%. The core of DSpark's technology lies in its semi-autoregressive architecture and the introduction of a confidence-based scheduling verification mechanism. This design aims to balance the speed of draft generation with the coherence of the final output. Furthermore, the framework features dynamic adaptability, allowing it to flexibly adjust verification length based on current system load, thereby achieving efficient inference acceleration under various pressure scenarios.

## Sources

- [APPSO](https://mp.weixin.qq.com/s/DRLYw-nSwquotCnlfNBpvg) (2026-06-29)

This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.