DeepSeek-V4-Flash Officially Released: AI Agent Capabilities Level Up! There's good news for developers and AI technology enthusiasts.

As of July 31, 2026, the DeepSeek-V4-Flash API officially enters the public beta phase. This update focuses on quite significant performance improvements, especially for AI agents and execution of complex technical tasks.
For those of you who frequently use the DeepSeek API, there is no need to worry about reconfiguration. The calling method has not changed at all. You just need to change the model name to deepseek-v4-flash in your application code to immediately enjoy this latest version.

Drastically Improved AI Agent Performance

One of the main advantages of this update is the surge in agentic capabilities. In various benchmark trials, DeepSeek-V4-Flash shows results that even surpass the V4-Pro-Preview version.
Here are some official benchmark scores achieved:

• 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 (Internal test for full-stack development)
• DSBench-Hard: 59.6 (Internal test for high-level coding problems)

Testing Note: For Code Agent tasks on public benchmarks, the DeepSeek team tested using the "DeepSeek Harness minimal mode" framework (to be released soon) with max effort level settings, top_p = 0.95, and temperature = 1.0.

Support for New Formats and Codex Integration

This official DeepSeek-V4-Flash version already supports the Responses API format by default (native). In addition, this model has been specifically adjusted to be compatible and optimal when used with Codex. For complete technical setup instructions, you can directly check the available official documentation.

Same Architecture, Better Optimization

Some of you might be curious whether the model size has become heavier? The answer is no.
This DeepSeek-V4-Flash-0731 model still uses exactly the same architecture and parameter size as the DeepSeek-V4-Flash-Preview version. The main difference lies in the re-post-training process carried out by the development team to polish its intelligence and response efficiency.

Important Things to Note

So that you are not confused, here are some important points regarding the scope of this update:

1. This update only applies to the DeepSeek-V4-Flash API.
2. The DeepSeek-V4-Pro API as well as models in mobile and web applications remain unchanged for now.
3. The official release for DeepSeek-V4-Pro will follow shortly.
   For those of you who want to try the robustness of the AI agent for coding, automation, or system analysis, now is the right time to try the deepseek-v4-flash API.