DeepSeek has officially launched the public beta of the API for its lightweight V4-Flash model. The results of internal testing were unexpected: the new version not only significantly enhanced agentic capabilities but also outperformed the heavier and more expensive V4-Pro-Preview on several metrics.
What changed in the architecture
The update only affected the API of the V4-Flash model. The senior V4-Pro, as well as the official app and web version of DeepSeek, remain unchanged for now. The key focus is on agentic tasks—scenarios where the model acts independently: writing and running code, searching for information, and executing chains of actions.
V4-Flash now directly supports the Responses API format—an interface through which an agent exchanges requests with the model and receives results. Additionally, the model has been adapted for Codex, OpenAI's popular coding agent. This allows developers to integrate DeepSeek into their familiar tools without changing their working environment. The API calling method remains unchanged—simply specify the model name deepseek-v4-flash in the settings.
The architecture and model size remain the same. The V4-Flash-0731 version retains the structure of the previous release but underwent additional retraining, which delivered this noticeable performance boost.
Test results: a breakthrough in agentic tasks
The new version outperformed both the previous V4-Flash-Preview and the older V4-Pro-Preview almost everywhere, at times approaching Anthropic's flagship Opus-4.8 and surpassing China's GLM-5.2.
In the Terminal Bench 2.1 test, which evaluates the ability to perform command-line tasks, the model scored 82.7 points, compared to 61.8 for the previous version. This is higher than GLM-5.2 (81.0) and nearly on par with Opus-4.8 (85.0).
The progress in programming is especially impressive. In NL2Repo (converting a text description into a ready-made repository), V4-Flash scored 54.2 points, up from 39.4 previously, and in DeepSWE (solving real-world programming tasks), it scored 54.4, up from a modest 7.3 for the previous version. A sharp jump is also visible in security: in Cybergym (finding vulnerabilities in code), the model scored 76.7 points, compared to 38.7 for its predecessor and 52.7 for the senior V4-Pro-Preview.
Results with tool use also improved. In Toolathlon-Verified (using external tools), V4-Flash scored 70.3 points, in Agents' Last Exam (complex agentic tasks)—25.2, and in AutomationBench (automating work scenarios)—25.1, surpassing previous versions across all three metrics.
The company's internal tests also showed growth: in DSBench-FullStack (full-fledged web development), the model scored 68.7 points, and in DSBench-Hard (complex tasks for a software agent)—59.6, notably higher than previous results, although still behind Opus-4.8.
What this means for developers
The public beta means open access to the API. Any developer can integrate the new version right now and test it on their tasks. The company has maintained backward compatibility: previous API calls continue to work, and switching to V4-Flash is as simple as changing the model name.
The update to the senior V4-Pro has been announced separately—its official release will come later.
My take: This move by DeepSeek is a serious signal for the AI market. The fact that a lightweight and cost-effective model outperforms the senior version of the previous generation speaks to rapid progress in architecture optimization. For developers, this means access to top-tier model performance at significantly lower infrastructure costs, which could accelerate the adoption of agentic solutions in mass-market products.