DeepSeek has launched the official public beta of the DeepSeek-V4-Flash API, featuring significantly enhanced agent capabilities compared to the V4-Pro-Preview. The update involves re-post-training the model while maintaining its original architecture and size.

  • Terminal Bench 2.1: 82.7
  • NL2Repo: 54.2
  • Cybergym: 76.7
  • DeepSWE: 54.4
  • Toolathlon verified: 70.3
  • Agent Last Exam: 25.2
  • Automation Bench (Public): 25.1
  • DSBench-FullStack: 68.7
  • DSBench-Hard: 59.6

The API natively supports the Responses API format and is specifically adapted for Codex, with benchmark results tested using the DeepSeek Harness minimal mode framework.