Frontline Hotspot
Frontline Hotspot

DeepSeek-V4-Flash Official API Public Beta: Agent Benchmarks Far Exceed V4-Pro-Preview

On 2026-07-31 DeepSeek launched the official (stable) V4-Flash API to public beta; the model name stays deepseek-v4-flash, with the same architecture as Preview, only re-post-trained. Agent capability is greatly enhanced, with official benchmarks far exceeding V4-Pro-Preview (Terminal Bench 2.1 82.7, Cybergym 76.7, DeepSWE 54.4, etc.). It natively supports the Responses API and is adapted for Codex; the V4-Pro official version is coming next.

Published July 31, 20266 min read
<!-- deepseek-v4-flash-hotspot | hotspot | DeepSeek-V4-Flash Official API Public Beta: Agent Benchmarks Far Exceed V4-Pro-Preview -->

On July 31, 2026, DeepSeek pushed the official (stable) DeepSeek-V4-Flash API into public beta. The call interface is unchanged; the model name is still deepseek-v4-flash, now pointing at the latest 0731 official build. The real headline isn't "another API shipped"-it's that agent capability is greatly enhanced, with official benchmarks far exceeding V4-Pro-Preview. The architecture and size are identical to April's Preview (284B total / 13B active MoE); only post-training was redone. V4-Pro's official version didn't make this round-official says "coming soon."

1. The public beta: what changed and what didn't

First, only V4-Flash's API was upgraded; V4-Pro's API and the App/Web models are unchanged this round, with V4-Pro official coming soon. Second, the model name stays deepseek-v4-flash, the API call format is unchanged-just connect and you get the latest. Third, DeepSeek-V4-Flash-0731's structure and size are identical to April's Preview, only re-post-trained-so this isn't a bigger model, it's the same skeleton with agent capability trained up. Fourth, the official V4-Flash natively supports the Responses API format and is specifically adapted for Codex.

2. Agent gains: benchmarks far exceed V4-Pro-Preview

This is the core. In official benchmarks, the official V4-Flash's agent scores far exceed V4-Pro-Preview:

  • 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
  • DSBench-Hard: 59.6

Two calibers matter. First, for Code Agent tasks on public benchmarks, the official V4-Flash was tested using DeepSeek Harness minimal mode (coming soon) as the framework, at max tier, topp=0.95, temperature=1.0-these scores are under a specific framework and params, not a raw zero-shot model. Second, DSBench-FullStack and DSBench-Hard are DeepSeek's internal test sets (full-stack dev and hard Coding Agent respectively), not independently reproducible like Terminal Bench / Cybergym. Separate "public" from "internal" when reading these-don't compare internal scores against public ones directly.

3. Why "same structure, re-post-trained" is the signal

Same architecture and size, only re-post-trained, means the gain isn't from more parameters but from post-training (RL, tool-use alignment) squeezing more agent capability out of the same skeleton. For everyone, that's a signal: on agent tasks, post-training may beat upgrading to a bigger model. And since the structure is unchanged, self-hosters' hardware bar is the same as Preview (13B active is relatively friendly to consumer GPUs)-no "the official version got heavier and won't run" problem.

4. Practical impact for developers and everyone else

For developers, the action is to treat the deepseek-v4-flash-pointed version as official (model name unchanged) and re-test on agent tasks. If your V4-Flash runs Code Agent or long tool-call chains, this lift is most relevant; for plain classification/extraction, the感知 is modest. Codex users get the tailored adaptation per the docs. Note V4-Pro didn't upgrade this round-hard reasoning stays on V4-Pro-Preview. For everyone else, V4-Flash has no consumer surface-it's still API + open weights (MIT, self-hostable). Self-hosters' hardware bar is unchanged; on the API, agent tasks with thinking + max and a saturated cache (cache hit $0.0028) remains the cheapest play.

An "official launch" sounds routine, but the real signal here is: same architecture, only re-post-trained, and it pushes agent benchmarks far past the previous stronger V4-Pro-Preview. It boils down to one thing-on your most agent-round-trip-heavy task, can the official Flash err less and loop less than Preview? If yes, the upgrade is worth it. V4-Pro official is still on the way; hard-reasoning folks can wait.


References

This article is AI-assisted and human-edited. Last updated: 2026-07-31

Related

Frontline Hotspot

AI Weekly 004: Seven Releases in Seven Days, but the Real Signals Are Agents, Compliance, and Cost

This week (Jul 27-Aug 2) the AI world shipped seven releases, but three signals matter more: DeepSeek-V4-Flash's post-training pushed DeepSWE from 7.3 to 54.4 (hands-on 30/30, cost under 5 fen) and Kimi K3 topped coding leaderboards; the EU AI Act August 2 deadline landed (fines up to 7% of global turnover, extraterritorial); prefix cache hits at 0.02 yuan vs 1 yuan misses make cost engineering a new skill.

Aug 2, 20265 min read
Frontline Hotspot

AI Weekly 003: GPT-5.6 Restricted, DeepSeek Open-Sources Inference Acceleration, Agents Shift from Chat to Work

This week's hard signals: OpenAI GPT-5.6 restricted by US regulators + self-developed Jalapeño chip, DeepSeek open-sources inference acceleration framework (A100 tasks moved to consumer GPUs, latency down 40%), Anthropic context-engineering guide, Xinliu Yuansu M-FLOW rewrites agent memory. Domestic AI carves a different track on efficiency/open-source/landing.

Jul 25, 20264 min read