DeepSeek V4 Flash Scores 54.4 on DeepSWE, Surpassing Previous 8-Point Record

DeepSeek V4-Flash achieves 54.4 on DeepSWE benchmark, nearing Claude Opus 4.8 and far exceeding prior V4-Pro scores, though framework differences apply.

Woofun AI data shows that DeepSeek-V4-Flash attained a score of 54.4 on the DeepSWE long-range software engineering benchmark. This result approaches Claude Opus 4.8’s 59 points and significantly exceeds the previous V4-Pro-Preview score of 8. While trailing Kimi K3’s 69 points, it outperforms GLM-5.2’s 44 points, securing a top-tier position among domestic models.

The evaluation utilizes DeepSeek’s unreleased Harness minimalist mode, meaning performance reflects both model capability and execution framework rather than raw model metrics under identical conditions.