According to DeepSeek on X, the company open sourced its in-house inference framework DeepInfer, which powers DeepSeek-V4.1-Flash at scale to cut compute costs and latency.