DeepSeek has released an updated version of its V4-Flash model, delivering what the company describes as a substantial improvement in agent capabilities. The public beta, dubbed V4-Flash 0731, shows remarkable gains across multiple agent-focused evaluation frameworks.
According to benchmarks published on DeepSeek’s API documentation, the model achieved an 82.7 score on Terminal Bench, representing a jump of 25.8 points from the July preview version. The model also scored 70.3 on Toolathlon, gaining 18.5 points, and achieved first-time scores of 76.7 on Cybergym, 54.4 on DeepSWE, and 54.2 on NL2Repo.
Perhaps most notably, V4-Flash 0731 exceeded the performance of V4-Pro-Preview on several agent tasks, despite being positioned as a more accessible model variant. This suggests that the retraining-focused update—described by DeepSeek as an architecture-unchanged refresh—has successfully optimized the model for autonomous workflow execution.
The improvements appear to address one of the key challenges facing large language models in production: the gap between benchmark performance and real-world agent task completion. Terminal Bench and Toolathlon both evaluate models on multi-step reasoning and tool use scenarios that more closely mirror actual deployed AI agent workflows.
DeepSeek has made V4-Flash available through its public beta API, allowing developers to test the model immediately. The company noted that V4-Pro API and DeepSeek’s consumer application endpoints remain unaffected by the update, suggesting a staged rollout strategy for the different model variants.
Industry observers have noted that the timing of the release puts pressure on other Chinese AI labs to demonstrate comparable agent capabilities. The V4-Flash gains follow a period of intense competition among AI providers to deliver models excelling at autonomous task completion, a capability increasingly demanded by enterprise customers building AI agent systems.
The benchmark results also highlight the pace of improvement possible through targeted retraining, even without architectural changes. As foundation models approach the limits of pure scale, such optimization strategies may become increasingly important for delivering practical agent performance gains.