Chinese company DeepSeek has released a public beta version of an updated API for its lightweight model V4-Flash. And this is not just a routine update: test results demonstrate that the junior model has not only caught up with, but in a number of key metrics surpassed the senior V4-Pro-Preview. This represents a significant step forward in the field of agentic computing and autonomous task execution.

What has changed in V4-Flash

The main focus of the update is enhancing the model's agentic capabilities. Now V4-Flash is capable of independently writing and running code, searching for information on the internet, and building complex chains of actions without human involvement. A key innovation is support for the Responses API format, which allows an agent to directly exchange requests with the model and receive results. Additionally, the model has been adapted for Codex from OpenAI — a popular tool for writing code — which expands its applicability for developers. The API calling method has not changed: to switch to the new version, simply specify the model name deepseek-v4-flash in the settings.

The architecture and size of the model remain the same, but the V4-Flash-0731 version has undergone additional retraining. This has made it possible to achieve impressive results without changing the underlying structure.

Test results: the numbers speak for themselves

The updated V4-Flash-0731 outperformed not only the previous V4-Flash-Preview in almost all benchmarks, but also the senior V4-Pro-Preview. In the Terminal Bench 2.1 test, which evaluates command-line skills, the model scored 82.7 points versus 61.8 for its predecessor. This is higher than GLM-5.2 (81.0) and nearly matches Opus-4.8 from Anthropic (85.0).

Progress is especially noticeable in code-writing tasks. In the NL2Repo test, V4-Flash scored 54.2 points versus 39.4 for the previous version, and in DeepSWE — 54.4 versus a modest 7.3. The growth in the security domain is also impressive: in the Cybergym test, the model scored 76.7 points, while the previous version showed only 38.7, and the senior V4-Pro-Preview — 52.7.

Results in tool use have also improved: 70.3 points in Toolathlon-Verified, 25.2 in Agents' Last Exam, and 25.1 in AutomationBench. Across all three metrics, the new version surpasses the previous ones. DeepSeek's internal tests also confirm the trend: in DSBench-FullStack, the model scored 68.7, and in DSBench-Hard — 59.6, significantly higher than previous versions.

What this means for developers

V4-Flash is positioned as a fast and cost-effective option for large volumes of work. Codex support allows the model to be used in a familiar environment without changing tools. The public beta means open API access — any developer can connect the new version right now. Backward compatibility is preserved: old requests continue to work, and the transition comes down to simply changing the model name.

An update to the senior model V4-Pro has been announced separately, and its official release is expected later. However, it is already clear: DeepSeek is betting that efficiency and speed can be more important than "raw" power.

My take: This is not just a technical update, but a signal of a paradigm shift in the AI industry. If a lightweight model outperforms the senior one in agentic tasks, then optimization and retraining can be more important than simply scaling up parameters. For the market, this means competition is shifting toward practical efficiency, not just raw benchmarks. Developers should take a close look at V4-Flash — savings on computational resources at this level of quality could become a decisive factor.