DeepSeek-V4-Flash Stable API Enters Public Beta With Major Agent Improvements

DeepSeek has launched the stable V4-Flash API in public beta, reporting substantial gains on agent benchmarks, including 82.7 on Terminal Bench 2.1 and 70.3 on Toolathlon Verified.

Published 8/6/2026, 9:21:25 AMLast updated 8/6/2026, 9:21:25 AMEditorial publication

According to DeepSeek’s update log, the stable DeepSeek-V4-Flash API is now available in public beta. The calling method remains unchanged: developers can select deepseek-v4-flash to use the latest version. The release supports the Responses API and includes adaptations for coding workflows such as Codex.\n\nDeepSeek says the 0731 release retains the preview model’s architecture and scale, with the main change being additional post-training. Reported scores include 82.7 on Terminal Bench 2.1, 54.2 on NL2Repo, 76.7 on Cybergym, 54.4 on DeepSWE, and 70.3 on Toolathlon Verified.\n\nThe comparisons with V4-Pro Preview should be treated cautiously because some evaluations used DeepSeek’s own harness, fixed settings, or different benchmark versions. The supplied report also lists DSBench-FullStack inconsistently as both 6 and 68.7; the official update log should be used to resolve that discrepancy.

Source evidence 6

Source review does not guarantee permanent accuracy. Consider the original source, observation time, and later corrections.

DeepSeek V4正式版来了，新能力浮出水面，性价比之王开战h5.ifeng.com · supporting

# DeepSeek V4正式版来了，新能力浮出水面，性价比之王开战摘要：性价比之王要打 Agent 仗了凤凰网科技出品作者｜Dale 编辑｜董雨晴 7 月 31 日午后，凤凰网科技查询 DeepSeek 官网发现，DeepSeek-V4-Flash 正式版 API 上线公测。与以往 “Pro 强、Flash 弱”的分层逻辑不同，这次更新释放了一个值得关注的信号 ——Flash 正式版在多项 Agent 基准测试中的表现，已经逼近甚至超越了三个月前 V4-Pro 预览版的水平。图片官方更新日志显示，V4-Flash 正式版在 Terminal Bench 2.1 上拿到 82.7 分，NL2Repo 54.2 分，Cybergym 76.7 分，Toolathlon verified 70.3 分。而 V4-Pro 预览版在 Terminal Bench 2.0 上的得分为 67.9 分。需要说明的是，Terminal Bench 2.0 与 2.1 并非同一版本的测试集，直接对比并不完全公平。但一个激活参数仅 130 亿的轻量版，在 Agent 能力上跑出这样的分数，已经在说明后训练阶段的优化空间，可能比单纯堆参数更具杠杆效应。此外，DeepSeek也特别说明目前公测仅限API，App和网页端暂无法体验最新能力。DeepSeek-V4-Pro正式版将尽快发布。 “仅重新进行了后训练” DeepSeek 官方在更新日志中表示，“DeepSeek-V4-Flash-0731 的模型结构、尺寸和 DeepSeek-V4-Flash-preview 保持一致，仅重新进行了后训练。” [...] 生态层面的一次正面交锋如果说 2023 年的大模型竞争是 “拼通用能力”，2024 年是 “拼长上下文”，那么 2026 年的关键词，毫无疑问是 Agent。 Agent 能力 —— 即模型自主规划、调用工具、执行复杂任务的能力 —— 正在成为衡量大模型实力的新标尺。从 Terminal Bench（终端操作）、NL2Repo（代码仓库生成）到 Cybergym（网络安全任务）、SWE-bench（软件工程），一系列 Agent 基准测试正在重新定义什么是好模型。从全球范围看，Agent 能力的第一梯队目前仍由闭源巨头把持。据 b