【#Tech24H】Peking University, in collaboration with DeepSeek, has developed and open-sourced DSpark, an inference acceleration framework for large models. This technology effectively solves the inference efficiency bottlenecks of large language models under high-concurrency scenarios, boosting single-user text generation speed by 60% to 85% at the same throughput. The corresponding papers and source code are publicly released on GitHub. Test data indicates that DSpark achieves better single-round effective generation length than the two mainstream baselines Eagle3 and DFlash. In the case of Qwen3-4B, the performance is improved by 30.9% compared with Eagle3 and 16.3% compared with DFlash. It maintains the first-token latency advantage of parallel architectures and mitigates the decay of candidate validity for long sequences.