Google 在 2026 年 7 月 1 日的官方博客上,把 6 月这一整月的 AI 更新做了一次长篇总结(**”The latest AI news we announced in June 2026”**)。如果只看标题党,多数人会觉得”又是一次月度例行通告”;但真正动手把全文拆开后会发现,Google 这个月悄悄推了三条彼此独立、但又能互相借力的产品线:
1 2 3 4 5
本地化 Gemma 4 12B 16GB 内存即可跑的视觉+语音统一模型 Agent 能力 Gemini 3.5 Flash 桌面/移动/浏览器通用的 computer use 多模态 API Gemini Omni Flash 视频工作流原生多模态(公开预览) Live 翻译 Gemini 3.5 Live Translate 70+ 语言保留说话人语调 移动端平台 Android 17 浮动应用、生物识别锁机、Pixel Drop
Gemma 4 12B brings smart AI agents directly to your laptop. It runs locally using just 16GB of memory, combining a novel unified architecture with vision and native voice processing in a single streamlined system.
6 月这条更新里最容易写软文的,是 Gemini 3.5 Flash 的 computer use。但 Google 在博客里把口径压得相当克制:
We integrated computer use into Gemini 3.5 Flash, allowing you to build custom agents that can see, reason and take action across desktop, mobile and browser environments. The update improves performance for long-horizon and enterprise automation tasks, like continuous software testing and knowledge work.
工程团队需要注意三个关键词:
1 2 3 4 5 6 7 8 9 10
across desktop, mobile and browser -> 不是只接管浏览器(那是过去一年大多数 agent demo 的卖点) -> 桌面、App、网页三端统一
第三条线:多模态 API——Nano Banana 2 Lite + Gemini Omni Flash 公开预览
第三组更新是 API 层的两条多模态产品:
1 2 3 4 5 6 7 8
Nano Banana 2 Lite -> Google 称其为"最快、最具性价比的 Gemini Image 模型" -> 适合做图像实验、原型期快速迭代
Gemini Omni Flash(公开预览) -> "natively multimodal model for enterprises and developers" -> 第一个 Google 把视频工作流原生放进 API 入口的型号 -> 关键词:"custom, dynamic video workflows"
对一家已经在线上跑图像/视频生成 API 的团队,这一条最直接的影响是计费模型重构。过去用 Imagen 做图 + Veo 做视频,分别两条账单;现在 Gemini Omni Flash 把”图、视频、文本”统一到一个端点,未来很可能出现”一个产品 = 一个 API Key + 一个成本维度”的简化结构。
博客在描述 Gemini Omni Flash 时还藏了一句:
introducing a natively multimodal model for enterprises and developers to build custom, dynamic video workflows for the very first time.
“for the very first time” 这种措辞出现在 Google 官方 blog 里,意味着这是他们的旗舰级发布,而不是例行升级。对工程团队,值得把这一条直接抄进下个季度的技术雷达——尤其是视频相关业务。
第四条线:Live 翻译——Gemini 3.5 Live Translate 把 70 种语言留住了语气
这条对多数中国工程团队”看起来无关”,但博客原文值得展开看:
This new audio model for live speech-to-speech translation automatically detects more than 70 languages while preserving the speaker’s natural intonation and eliminating awkward pauses.
# 1. 强合规 + 能力匹配 -> 本地 if signal.privacy_required: ep = ENDPOINTS["gemma4_local"] if ( (not signal.has_screenshot or ep.supports_vision) and (not signal.has_voice_input or ep.supports_voice) ): candidates.append((0.0, ep.avg_step_latency_s, ep))
# 2. computer use 必需 -> Gemini 3.5 Flash if signal.needs_desktop_action: ep = ENDPOINTS["gemini_flash_cloud"] candidates.append((ep.cost_per_1m_tokens_usd, ep.avg_step_latency_s, ep))
# 3. 多模态兜底 if signal.has_screenshot or signal.has_voice_input: for key in ("gemma4_local", "gemini_omni_cloud"): ep = ENDPOINTS[key] if ( (not signal.has_screenshot or ep.supports_vision) and (not signal.has_voice_input or ep.supports_voice) ): candidates.append((ep.cost_per_1m_tokens_usd, ep.avg_step_latency_s, ep))
# 4. 没有特定要求:返回最便宜的 ifnot candidates: candidates = [ (ep.cost_per_1m_tokens_usd, ep.avg_step_latency_s, ep) for ep in ENDPOINTS.values() ]
candidates = [ c for c in candidates if c[0] <= signal.max_budget_usd ] ifnot candidates: returnNone candidates.sort(key=lambda x: (x[0], x[1])) # 优先 cost,再 latency return candidates[0][2]