ModelsMulti-source86°

Google 两周密集发布:Gemini 4 三款新模型、Genie 4、Robotics 2.5 一同亮相

Why we picked it

Google 这波一次发了 Gemini 4 的三档模型加 Genie 4 世界模型,做编码的看 Hydrogen,跑长任务的看 Carbon,Genie 4 还首次进了 API 和 AIStudio,搞机器人和生信的也有新东西,值得对照选型。

Google 在一条 Recap 中集中公布了近两周的多项发布。Gemini 4 系列包含三款模型:Neon 主打前沿智能与复杂软件工程、知识工作和科学工作流的快速端到端任务完成;Hydrogen 面向日常编码与智能体工作流的速度和效率;Carbon 则针对长周期项目和深度研究。

Body

Genie 4 是 GoogleDeepMind 称其最逼真的交互式世界模型,支持实时场景编辑和自适应原生音频,已上线 Project Genie(新增创作智能体和画廊),并首次通过 Gemini API 和 GoogleAIStudio 开放。Gemini 3.5 Transcribe 新增原生语音编辑,开发者可让用户用语音一步完成文本的编辑、改写和翻译。Gemini Robotics 2.5 是 DeepMind 迄今最强的视觉-语言-动作(VLA)模型,AlphaGenome 2 则聚焦特定细胞帮助科学家理解疾病驱动因素。

产品侧,GoogleLabs 的 Stitch Loop 实验会监控生产数据并持续循环改进设计产品直至目标达成。这一波发布覆盖语言模型、世界模型、语音、机器人和基因组学多条线,Gemini 4 的三档划分(前沿/效率/长任务)明显是针对不同工作负载做产品分层,与 Anthropic、OpenAI 的多档模型策略相似,开发者可在 AIStudio 中直接调用 Genie 4 和语音编辑能力。

Original · Google AI

We’ve been busy shipping 🚢 Here’s a recap of the last couple of weeks: — Gemini 4 Neon, our newest model that brings frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows — Gemini 4 Hydrogen, our model for speed and efficiency in everyday coding and agentic workflows — Gemini 4 Carbon, our model for your hardest long-horizon projects and deep research — Genie 4, @GoogleDeepMind ’s most realistic interactive world model featuring real-time scene editing and adaptive native audio, now available in Project Genie (with a new creation agent and gallery) and for the first time in the Gemini API and @GoogleAIStudio — Gemini 3.5 Transcribe, now with native voice editing that lets developers build experiences where users can edit, rephrase, and translate existing text in a single step using just their voice, via the Gemini API in @GoogleAIStudio — Gemini Robotics 2.5, @GoogleDeepMind ’s most capable vision-language-action (VLA) model to date — AlphaGenome 2, @GoogleDeepMind ’s new AI model that zooms in on specific cells to help scientists understand what drives disease — Stitch Loop, a new @stitchbygoogle experiment from @GoogleLabs that monitors production data and runs continuous loops to improve your designed product until goals are met📊 32 ⚡

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Gemini 4, Genie 4, Gemini Robotics, GoogleAIStudio.

Source attribution: curated from Google AI (@GoogleAI). Cross-source corroboration: 11 independent sources have covered this development (including OpenRouter, lmarena.ai, 歸藏(guizang.ai)), confirming industry-wide signal strength.