Chinese AI startup DeepSeek says it has introduced DSpark, an upgrade to its flagship V4 model designed to reduce inference bottlenecks and improve generation speed. Multiple outlets report that DSpark uses a technique DeepSeek describes as speculative decoding. The company claims the framework can increase per-user response speeds by up to 85%, indicating improved efficiency during model serving. The reports frame the upgrade as part of a broader shift in competition among Chinese AI developers, where attention increasingly focuses on lowering the cost of running models and improving user experience rather than only model quality. While the outlets highlight performance gains and the potential to reduce reliance on larger compute resources, they do not provide detailed technical specifications, benchmarking methodology, or independent verification in the provided excerpts. Overall, the information centers on DeepSeek’s announcement of DSpark and its claimed effect on faster AI responses when deployed with the V4 model.