[Peking University and DeepSeek Jointly Open Source DSpark Framework, Boosting Generation Speed by Over 60% Under High Concurrency] According to a report by Jinse Finance, on June 27, DeepSeek, in collaboration with Peking University, officially released the DSpark inference acceleration framework, aimed at addressing the inference efficiency bottleneck of large language models in high-concurrency production environments. The framework has been deployed in the preview service engines of DeepSeek-V4-Flash and DeepSeek-V4-Pro. Compared to the single-token prediction decoding baseline MTP-1 previously used in production environments, it can increase single-user generation speed by 60% to 85% under the same throughput level. Relevant papers, training codes, and model checkpoints have been open-sourced on the GitHub platform under the DeepSpec project.