DeepSeek has officially launched the public preview of the DeepSeek-V4-Flash API, marking another major milestone in the company’s rapidly evolving AI platform. The announcement was made on July 31 through DeepSeek’s official API documentation, where the company also confirmed that the official release of DeepSeek-V4-Pro is expected to arrive “as soon as possible.”
The update focuses on significantly enhancing the model’s AI agent capabilities, with DeepSeek claiming that the new V4-Flash outperforms the previously released V4-Pro Preview across multiple industry-standard benchmarks.
Stronger AI Agent Performance
According to DeepSeek, the latest V4-Flash has undergone extensive post-training optimizations that substantially improve its ability to perform complex, tool-using, and software engineering tasks.
The company published benchmark results demonstrating the model’s performance across a variety of AI agent evaluations:
| Benchmark | Score |
|---|---|
| Terminal Bench 2.1 | 82.7 |
| NL2Repo | 54.2 |
| Cybergym | 76.7 |
| DeepSWE | 54.4 |
| Toolathlon | 70.3 |
| Agent Last Exam | 25.2 |
| Automation Bench (Public) | 25.1 |
| DSBench-FullStack | 68.7 |
| DSBench-Hard | 59.6 |
DeepSeek states that these results place the official V4-Flash ahead of the earlier V4-Pro Preview in agent-oriented workloads, particularly those involving coding, automation, software engineering, and tool usage.
Native Responses API Support
One of the key improvements in the official release is native support for the Responses API format, allowing developers to integrate the model more easily into modern AI applications.
DeepSeek also notes that the model has been specifically optimized for OpenAI Codex-compatible workflows, making it suitable for coding assistants, software development platforms, and autonomous programming agents.
The company has published updated configuration documentation to help developers migrate existing projects and begin testing the new API.
Same Model Architecture, Improved Training
Despite the performance improvements, DeepSeek clarified that DeepSeek-V4-Flash-0731 retains the same underlying architecture and parameter scale as the earlier DeepSeek-V4-Flash Preview.
Rather than introducing a larger model, the company focused on retraining the post-training stage, enabling better reasoning, stronger agent behavior, and improved task execution without changing the core model structure.
This approach allows developers already using the preview version to benefit from higher performance while maintaining compatibility with existing implementations.
V4-Pro Remains Unchanged—for Now
DeepSeek emphasized that this update applies only to the V4-Flash API.
At present:
- DeepSeek-V4-Flash API has been upgraded to the official public preview version.
- DeepSeek-V4-Pro API remains unchanged.
- DeepSeek’s web platform and mobile applications continue to use the existing production models.
The company confirmed that the official version of DeepSeek-V4-Pro is currently in development and will be released “as soon as possible,” although no specific launch date has been announced.
Expanding DeepSeek’s AI Ecosystem
The release of the official DeepSeek-V4-Flash API underscores the company’s continued focus on AI agents, software engineering, and developer tools. By improving benchmark performance while maintaining architectural compatibility, DeepSeek aims to provide developers with a more capable and production-ready model for building next-generation AI applications.
With V4-Pro still on the horizon, today’s release offers an early look at the direction of DeepSeek’s latest generation of large language models and signals the company’s continued investment in high-performance AI systems for enterprise and developer use.