[Old Pro only scored 8 points, DeepSeek-V4-Flash reached 54.4 points on DeepSWE, approaching Opus 4.8] According to monitoring by Beating, on the long-range software engineering benchmark DeepSWE, the official version of DeepSeek-V4-Flash achieved a score of 54.4, approaching Claude Opus 4.8's 59 points, and significantly surpassing the previous 8 points of V4-Pro-Preview. This result is lower than Kimi K3's 69 points but higher than GLM-5.2's 44 points, placing it firmly in the top tier of domestic models. However, this is not a pure model test under identical conditions. The official version of V4-Flash utilized DeepSeek's yet-to-be-released Harness minimal mode, where Agent performance is influenced by both the model and the execution framework. [Original link]