Google DeepMind 发布 Gemini 3.8 Flash 与 Gemini 3.8 Flash Cyber：同一核心模型、两种访问方式

谷歌刚刚宣布了 Gemini 3.8 Flash 和 Gemini 3.8 Flash Cyber：https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/，在 Gemini 3.7 Flash 之后三周：https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-3-7-flash/，标志着六周内的第三次 Flash 发布。两种版本都运行在相同的基础智能之上，通过长期的代理循环对底层模型进行递归评估而提升。区别不在于架构，而在于安全防护范围及允许访问的人群。

它可以部署吗？Gemini 3.8 Flash 今天就可以通过 Gemini API 使用：https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash、Google AI Studio、Antigravity：https://antigravity.google/、Android Studio 和 Gemini Enterprise：https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-8-flash，所以你现在就可以将生产流量引向它。模型权重是封闭的，因此没有自托管或本地部署方式。Gemini 3.8 Flash Cyber 完全不能公开部署：访问权限通过新的 Fairwind 计划按案例授予：https://deepmind.google/fairwind-program/。

研究团队：https://deepmind.google/models/model-cards/gemini-3-8-flash/explains 解释说 3.8 Flash 是基于 3.7 Flash 的。规格保持不变：1048576 令牌上下文窗口，65536 令牌最大输出，文本、图像、音频和视频输入，以及文本输出。思维水平仍为 LOW、MEDIUM 和 HIGH，默认为 MEDIUM。对于迁移的用户，有一个重要变化：3.8 Flash 不支持 MINIMAL，设置该选项会返回 API 验证错误。

行为变化是重点。谷歌直白地描述了收益：3.8 Flash 工作更努力。在复杂任务中，它会执行额外的推理步骤并迭代调用工具，并且在高强度模式下可能消耗更多的令牌。谷歌自己的开发者指南：https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/guides/gemini-3-8-flash 坦率地指出，这样做以更高的令牌消耗换取更好的准确性，并建议当计算效率是约束条件时继续使用 3.7 Flash。这是一个不寻常的直接承认，即新模型并非每种工作负载的默认最佳选择。

在长远软件工程基准 DeepSWE v1.1 上，Google 报告称 3.8 Flash 在成本仅为部分费用的情况下，超越了大多数更大的前沿模型。它在 HLE-Verified 上记录了 54.9%，并且在 Vals Finance Agent V2：https://www.vals.ai/benchmarks/fabv2 以及 Harvey’s Legal Agent Benchmark：https://www.vals.ai/benchmarks/hlab 上，比 3.7 Flash 和其他前沿模型有更高的表现。注意，金融和法律结果是以相对胜利的形式报告的，公告中未提供绝对分数。

在标准漏洞发现基准 CyberGym 上，Google 报告前沿水平的性能，超越了 3.5 Flash Cyber 和明显更大的前沿模型，但没有发布绝对数值。由于 CyberGym 主要是 C 和 C++，Google 还运行了一个覆盖 20 种编程语言的内部基准，报告漏洞发现成功率超过 70%。

针对补丁，Collinear 运行的 CWE-Bench：https://cwe-bench.com/#leaderboard，将 Flash Cyber 的 pass@1 评为 47.2%，而领先的前沿模型为 47.8%。Google 称其在成本显著较低的情况下已接近平价，并将其定位为处于帕累托前沿，而非排行榜首位。

Chrome 安全报告称，Flash 生成的正确补丁比最好的大规模商业模型多 2.6 倍。Wiz：https://www.wiz.io/ 在其内部渗透测试基准上测量出 7.5 到 9.7 个百分点更高的召回率，同时成本降低 2.3 到 5.2 倍。Google 的云漏洞研究团队在不到两小时内发现了一个关键基础漏洞，而这项工作通常需要数月时间。

Google 明确表示，它优先考虑漏洞修复，而非利用等攻击性能力。Flash Cyber 附带了一套更宽松的网络缓解措施，这正是它被限制仅供受信任的防御者使用的原因：政府机构、关键基础设施运营商以及通过 Fairwind 申请的软件维护者。

需要与我们合作推广你的 GitHub 仓库 或 Hugging Face 页面 或 产品发布 或 网络研讨会等？请联系我们：https://forms.gle/wbash1wF6efRj8G58

Asif Razzaq 是 Marktechpost Media Inc. 的首席执行官。作为一位有远见的企业家和工程师，Asif 致力于利用人工智能的潜力造福社会。他最近的努力是推出人工智能媒体平台 Marktechpost，该平台因其对机器学习和深度学习新闻的深入报道而脱颖而出，既技术上可靠，又易于广大受众理解。该平台每月浏览量超过 200 万次，显示了其在观众中的受欢迎程度。

Google just announced Gemini 3.8 Flash and Gemini 3.8 Flash Cyber：https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/, three weeks after Gemini 3.7 Flash：https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-3-7-flash/ and marking the third Flash release in six weeks. Both variants run on the same foundational intelligence, refined through long-running agentic loops that recursively evaluate the underlying models. What separates them is not architecture, it is the safety envelope and who is allowed through it.

Is it deployable? Gemini 3.8 Flash is generally available today through the Gemini API：https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash, Google AI Studio, Antigravity：https://antigravity.google/, Android Studio, and Gemini Enterprise：https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-8-flash, so you can route production traffic to it now. Weights are closed, so there is no self-hosted or on-premises path. Gemini 3.8 Flash Cyber is not openly deployable at all: access is granted case by case through the new Fairwind Program：https://deepmind.google/fairwind-program/.

The research team ：https://deepmind.google/models/model-cards/gemini-3-8-flash/explains that 3.8 Flash is based on 3.7 Flash. Specs are unchanged: a 1,048,576-token context window, 65,536-token maximum output, text, image, audio, and video input, and text output. Thinking levels remain LOW , MEDIUM , and HIGH with MEDIUM as the default. One breaking detail for anyone migrating: MINIMAL is not supported on 3.8 Flash and setting it returns an API validation error.

The behavioral change is the point. Google describes the gain bluntly: 3.8 Flash works harder. On complex tasks it executes extra reasoning steps and calls tools iteratively, and it may burn more tokens at higher effort levels. Google’s own developer guide：https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/guides/gemini-3-8-flash is candid that this buys better accuracy at the cost of higher token consumption, and it recommends staying on 3.7 Flash when compute efficiency is the binding constraint. That is an unusually direct admission that the newer model is not the right default for every workload.

On DeepSWE v1.1, a long-horizon software engineering benchmark, Google reports 3.8 Flash outperforming most larger frontier models at a fraction of the cost. It records 54.9% on HLE-Verified, and gains over 3.7 Flash and other frontier models on Vals Finance Agent V2：https://www.vals.ai/benchmarks/fabv2 and Harvey’s Legal Agent Benchmark：https://www.vals.ai/benchmarks/hlab. Note that the finance and legal results are reported as relative wins, without absolute scores in the announcement.

On CyberGym, the standard vulnerability discovery benchmark, Google reports frontier-level performance that surpasses both 3.5 Flash Cyber and significantly larger frontier models, though no absolute figure is published. Because CyberGym is mostly C and C++, Google also ran an internal benchmark across 20 programming languages and reports a discovery success rate above 70%.

For patching, CWE-Bench：https://cwe-bench.com/#leaderboard, run by Collinear, puts Flash Cyber at 47.2% pass@1 against a leading frontier model’s 47.8%. Near-parity at materially lower cost is the claim, and Google frames it as sitting on the Pareto frontier rather than topping the leaderboard.

Chrome Security reports 2.6x more correct patches than the best, much larger commercial models. Wiz：https://www.wiz.io/ measures 7.5 to 9.7 percentage points higher recall on its internal penetration testing benchmark at 2.3x to 5.2x lower cost. Google’s Cloud Vulnerability Research team found a critical foundational vulnerability in under two hours, work that normally takes months.

Google is explicit that it prioritized vulnerability fixing over offensive capabilities like exploitation. Flash Cyber ships with a more permissive set of cyber mitigations, which is precisely why it is restricted to trusted defenders: government authorities, critical infrastructure operators, and software maintainers who apply through Fairwind.

Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us ：https://forms.gle/wbash1wF6efRj8G58

Asif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of Artificial Intelligence for social good. His most recent endeavor is the launch of an Artificial Intelligence Media Platform, Marktechpost, which stands out for its in-depth coverage of machine learning and deep learning news that is both technically sound and easily understandable by a wide audience. The platform boasts of over 2 million monthly views, illustrating its popularity among audiences.