终于，在吊了大家很久胃口之后，昨晚 Gemini 3 上线。用近乎恐怖的实力，碾压各大模型。 一句话就能生成 3D 模型、做网站，甚至做一个开放世界游戏…… 现在，进入 Google AI Studio，你就能直接体验 Gemini 3 Pro 预览版。至于面向更加大众的 Gemini 网站和 app，也会很快上线。 我不是针对谁，我是说在座的各位…… Gemini 3 Pro 晒出成绩单，它不仅完全把前辈 Gemini 2.5 Pro 拍死在沙滩上，还在除"解决真实 GitHub 问题（SWE-Bench Verified）"这一项之外，全面碾 压了 Claude Sonnet 4.5 和 GPT-5.1。 这就好像一个班里有几个语数外偏科的尖子生，这时候来了一个各科满分的三好生小霸王，你说气人不？吓人不？ | Benchmark | Description | | Gemini 3 Pro | Gemini 2.5 Pro | Claude Sonnet 4.5 | GPT-5.1 | | --- | --- | --- | --- | --- | --- | --- | | Humanity's Las ...

凌晨，谷歌终极杀器Gemini 3重磅来袭，一出手就是Pro顶配版，号称「史上最强推理+多模态+氛围编程」三合一AI战神！基准测试横扫全 场，就连GPT-5.1也被斩于马下，AI的下一个时代开启。 它来了，它来了！ 就在刚刚，万众期待的年度压轴之王，谷歌新一代旗舰Gemini 3炸裂登场。 而且，一上来就是顶配的Gemini 3 Pro—— 迄今推理最强，多模态理解最强，以及「智能体」+「氛围编程」最强的模型！ 强到什么程度？ 发布一小时后，就连OpenAI CEO奥特曼，都亲自发推表示祝贺！ 而且，还是区分大小写的版本。（不知道是不是亲自试了一下） 从实测来看，也的确如此。 在众多基准测试中，Gemini 3 Pro一举封神—— 不仅相较于2.5 Pro实现了性能的全方位跃升，甚至直接把OpenAI刚上新的GPT-5.1甩出了好几条街。 | | Rank 14 | Rank Spread O (Upper-Lower) | Model 11 | Score ↓ | | --- | --- | --- | --- | --- | | | 1 | 1 =- 2 | G gemini-3-pro | 1501 | ...

Model Performance & Benchmarks - Gemini 3 surpasses previous Frontier models in benchmarks, demonstrating significant advancements in AI capabilities [1] - Gemini 3 achieves 458% with code execution and search on Humanity's last exam, compared to Gemini 25% Pro at 21%, Cloud Sonnet 45% at 13%, and GBT 51% at 265% [2] - On the Vending Bench benchmark, Gemini 3's net worth reached $547816%, significantly outperforming Cloud Sonnet 45% at $3800 [4] - Gemini 3 Deep Think scores 41% on Humanity's Last Exam, compared to Gemini 3 Pro at 375%, Claude Sonnet 45% at 13%, GPT5 Pro at 30%, and GPT 51% at 265% [9][10] - Gemini 3 Deepthink achieves 451% on Arc AGI2 visual reasoning puzzles, a 10x improvement over Gemini 25% Pro [12] Enterprise Applications & Features - Boxcom's benchmark shows a 22-point performance increase for Gemini 3 Pro versus Gemini 25% Pro, with scores of 85% and 63% respectively [6] - Industry subsets in Boxcom's benchmark show significant performance jumps: Healthcare and Life Sciences (45% to 94%), Media and Entertainment (47% to 92%), and Financial Services (51% to 60%) [6] - Gemini 3 excels in complex multi-step reasoning and task automation, as highlighted by Box's new benchmark [7] - Gemini 3 supports multiple modalities, including text, images, video, audio, and code, with a unique focus on video understanding [12] - Gemini 3 can analyze YouTube videos frame by frame, understanding the content in detail [13] Google Integration & New Products - Gemini 3 is integrated into Google Search, dynamically generating user interfaces based on user queries [17] - Google launched anti-gravity, a VS Code fork coding platform that supports Gemini models and other models like GPTOSS and Anthropic's Sonnet [20] - The updated Gemini app features Gemini Agent capability, enabling the AI to complete real tasks on the user's behalf and create dynamic UIs [24] Model Architecture & Specifications - Gemini 3 is a brand new foundation model, not a modification of a prior model [27] - The model accepts text, images, audio, and video files as inputs, with a token context window of up to 1 million and output tokens of 64000 [28] - Gemini 3 is a sparse mixture of experts model built on Google's custom TPU architecture for both pre-training and inference [28]

RT Logan Kilpatrick (@OfficialLoganK)And say hello to Gemini 3 Deep Think, even more SOTA compared to Gemini 3 Pro 🤯 https://t.co/DJ2DePWJTd ...