DeepSeek Releases V4-Flash in Public Beta with a Focus on Agents

DeepSeek has opened the public beta of its updated DeepSeek-V4-Flash model. The way you call the API hasn't changed: you simply specify the model name deepseek-v4-flash. The update affects only the API version of Flash — the V4-Pro interfaces, as well as the models in the app and on the web, remain unchanged. An official V4-Pro release is promised soon.

Betting on agentic tasks

The main emphasis of the new version is its work as an agent. According to the developers, its results on the relevant benchmarks noticeably exceed those of the V4-Pro preview. Among the figures cited:

• 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

Internal test suites are mentioned separately: DSBench-FullStack (full-fledged web development) — 68.7 and DSBench-Hard (hard coding-agent tasks) — 59.6.

At the same time, the company is upfront about the conditions of the measurements: for the Code Agent tasks from the public suites, they used the DeepSeek Harness framework in minimal mode (which has yet to be released) with the maximum “effort” level, a top-p of 0.95, and a temperature of 1.0. In other words, reproducing these numbers “out of the box” isn't possible just yet.

Compatibility and how the model is built

The official V4-Flash natively supports the Responses API format and has been specially adapted for Codex — for the configuration, the developers refer you to the documentation. A curious detail: the DeepSeek-V4-Flash-0731 version keeps the same architecture and size as the preview — it was merely re-tuned during the post-training stage. In other words, the gain in quality was achieved without changing the model itself, through fine-tuning alone.

As a reminder, DeepSeek introduced the V4 family (Pro and Flash) in April 2026, making it available both through the OpenAI ChatCompletions interface and through the Anthropic API. The old model names deepseek-chat and deepseek-reasoner, which served as entry points to the non-thematic and “thinking” modes, were slated by the company to be retired once a three-month transition period elapsed.

The announcement drew the community's attention: on Hacker News the post gathered 205 votes and 84 comments.

Source: Hacker News