模型与实验室 4.0 · 优秀 2026-08-13 · 文章

Introducing Gemini 3.7 Flash

Google 发布 Gemini 3.7 Flash,距 3.6 Flash 仅三周,定位编码与智能体场景的主力工作模型FrontierCode 1.1 Main 43.6%(3.6 Flash 为 34.4%),DeepSWE v1.1 65.3% vs 49.0%,WebDev Arena Elo 1588 vs 1538,GDP.pdf 34.0% vs 22.0%,AutomationBench 30.4% vs 17.0%;引入价为 3.6 Flash 每百万 token 的一半收录理由:主力级模型的性价比与工作流基准同步大幅跃升,对 coding agent 选型有直接参考价值

打开原文回到归档

Introducing Gemini 3.7 Flash

中文导读

Google 发布 Gemini 3.7 Flash,距 3.6 Flash 仅三周,定位"编码与智能体场景最智能的主力工作模型"。编码与复杂工作流全面超越 3.6 Flash:FrontierCode 1.1 Main 43.6% vs 34.4%(生产级代码首过准确率)、DeepSWE v1.1 65.3% vs 49.0%(长程软件工程)、WebDev Arena Elo 1588 vs 1538(Web 开发,更少 prompt 生成功能完整的布局与应用,UI 生成对截图/图像/设计系统参考有高还原度)、GDP.pdf 34.0% vs 22.0%(金融/法律/生物等知识密集文档理解)、AutomationBench 30.4% vs 17.0%(真实业务工作流自动化)。开发者体验改进:更能适应阻碍、必要时澄清意图、指令遵循保真度更高,多步规划与工具调用更"勤勉",减少人工监督与重试。年底前的引入价为每百万 token 输入 $0.75 / 输出 $3.75——即 3.6 Flash 原价的一半。Gemini Spark(面向 AI Pro/Ultra 订阅者的 24/7 个人 agent)即日起升级到 3.7 Flash,Workspace 工具使用与多技能工作流精度提升。安全侧更新 CBRN 与网络攻击滥用的防护。

为什么值得关注

主力级模型的性价比与工作流基准同步大幅跃升,对 coding agent 选型有直接参考价值:半价 + 编码/自动化基准接近翻倍(AutomationBench 17.0→30.4),意味着生产级 agent 的单任务成本与可靠性同时改善。发布节奏(3 周迭代)与"developer feedback + algorithmic innovations 前移到未来模型"的说法,也预示 Flash 系列成为 Google 快速消化工程改进的载体。

关键信息

  • 发布:2026-08-13;作者 Tulsee Doshi(Google Gemini 团队)
  • 定位:most intelligent workhorse model yet for coding and agents
  • 编码:FrontierCode 1.1 Main 43.6%(3.6 Flash 34.4%);DeepSWE v1.1 65.3%(49.0%)
  • Web 开发:WebDev Arena Elo 1588(1538)
  • 文档:GDP.pdf 34.0%(22.0%)
  • 自动化:AutomationBench 30.4%(17.0%)
  • 价格:年底前引入价 $0.75/1M 输入 + $3.75/1M 输出(3.6 Flash 原价的 1/2)
  • 生态:Gemini Spark 即日升级;AI Studio / Antigravity / Android Studio / Gemini Enterprise 可用;模型卡见 deepmind.google
  • 安全:更新 CBRN 与 cyber offense 滥用防护

正文存档(要点摘录)

This release comes just three weeks after Gemini 3.6 Flash, and is a direct result of developer feedback and algorithmic innovations that we look forward to bringing to future models. 3.7 Flash delivers substantial improvements across software engineering, knowledge work, and web development workflows — with an introductory price of half the original 3.6 Flash cost per million tokens.

On developer experience: "3.7 Flash better adapts to roadblocks, clarifies intent when needed, and follows instructions with greater fidelity. It thinks more diligently, putting in more effort into multi-step planning and tool calls. A more disciplined execution means less manual oversight and fewer retries across engineering workflows."

Gemini Spark, the 24/7 personal agent for Google AI Pro and Ultra subscribers in 160+ countries, moves to 3.7 Flash starting today — more efficient for knowledge work with improved tool use for Google Workspace apps, delivering improved accuracy and output quality for complex, multi-skill workflows.

Demos in the post: text prompt → fully playable 3D game (with Nano Banana generating characters/items/textures in real time); single-shot interactive landing pages (3.7 Flash orchestrating sub-agents, Gemini Omni for parallax components); robotics model training with multimodal understanding in a 3-agent graph loop; static PDF → interactive data story with live charts.

Safety: updated safeguards against misuse in CBRN and cyber offense domains per Google's bioresilience approach and cyber program.

Obsidian Notes

  • 正文存档由 opencli web read 抓取 Google 官方博客全文后摘录整理;基准数字、价格与 Spark 升级信息均直接来自原文。
  • 相关条目:e6c53596(GLM-5.3,同期发布的开源权重编码模型)、afb6d0b8(Qwen3.8-27B-FP8)。三者均引用 DeepSWE v1.1,可横向对比:GLM-5.3 66.9 / Gemini 3.7 Flash 65.3 / Qwen3.8-27B 42.2。