Google 发布 Gemini 4 Hydrogen 模型,升级搜索复杂问题回答能力
Google’s latest Search upgrade promises better answers to your most complicated questions
- 1Google 发布 Gemini 4 Hydrogen 模型并接入搜索,官方称这是搜索迄今幅度最大的模型改进
- 2官方称 Hydrogen 完成任务速度比前代快模型 Gemini 3.8 Flash 快 2.5 倍,并在默认 high 档的 Artificial Analysis Intelligence Index 上超过 Gemini 4 Argon
- 3搜索侧改进覆盖多部分复杂问题、拼写错误容错和 fan-out 子查询拆分,AI Mode 将出现更多图片、结构化回答和对比表格
- 4模型提供 low、medium、high 和新增 xhigh 四档思考级别,将登陆 AI Studio、Gemini API、Gemini App 及 Antigravity 平台,并向全球搜索用户推送
Google 把搜索回答换到更快更便宜的模型档位,并用四档思考级别把模型强弱变成可调的成本与速度参数。
- 影响:搜索 AI Mode 的回答会更依赖图片、对比表格等结构化形式,网站被召回取决于子查询拆分的覆盖面;用 Gemini API 的企业多了一个速度优先、输出 token 更少的模型档位。
- 参考:可考虑在 Gemini API 上挑高并发轻量任务试用 Hydrogen,对比耗时与 token 成本;内容侧值得检查结构化数据和常见问题表述,适配 AI Mode 的对比表格与子查询召回。
- 依据:速度 2.5 倍、超越 Gemini 4 Argon 及 Agent Arena 第 5 名均为 Google 官方口径,暂无独立验证;推送范围与订阅档位以官方公告为准。
Google Search is getting better at understanding complicated questions, including those with multiple parts, highly specific requests, or confusing typos. The updates are coming via the company’s newly released Gemini 4 Hydrogen model, which Google claims is its “biggest model improvement to Search yet.”
While Google’s frontier models, Gemini 4 Carbon and Gemini 4 Argon, are best at deep reasoning for hard problems, Hydrogen is designed for fast execution across a variety of agentic workflows, or tasks in which an AI carries out multiple steps on its own, the company explained in an announcement.
The idea behind the model’s release is that using the most powerful model isn’t always needed — sometimes a faster, cheaper, and reasonably smart one will do. Google says Hydrogen completes tasks 2.5 times faster than Gemini 3.8 Flash, its previous fast model, but it’s still pretty smart, even beating the more powerful Gemini 4 Argon on one broad intelligence benchmark, the Artificial Analysis Intelligence Index, at its default high thinking level.
In Search, Hydrogen brings several improvements, including a better understanding of user intent, the ability to offer more helpful links and visual responses, and better answers across the different languages supported by Google’s AI Mode.
In terms of intent, Google says Hydrogen can handle more nuanced and complicated questions more quickly. This means people can ask their questions in their natural train of thought, including those that have multiple parts or that include highly specific requests. Hydrogen is also better at navigating hard-to-understand typos, Google says. Because of its improvements and better ability to follow instructions and call the right tools, Hydrogen is able to more precisely address each part of a user’s question.
Its advanced reasoning also improves Search’s fan-out technique, which breaks down a single user’s prompt into multiple subqueries. This allows people to discover websites related to the query that they may have otherwise missed, the blog post explains.
The addition of Hydrogen to Search will also mean more images, structured responses, comparison tables, and custom tools will appear as visual responses to users’ queries in AI Mode.




Google also says Hydrogen can work on its own for everyday work in areas like writing code, running tests, or verifying code changes, or it can be paired with Carbon and Argon in workflows that combine several models.
The model offers four thinking levels: low, medium, high (default), and a new, extra-high tier (“xhigh”), all of which are designed to complete tasks faster with fewer output tokens.
On multi-step tasks like deep research, data analysis, and broad codebase navigation, Hydrogen ranks No. 5 on Agent Arena’s benchmarks, the company pointed out.
In addition to Search, Gemini 4 Hydrogen is coming to all tiers of Google’s AI Studio, the Gemini API, the Gemini App, and the paid AI Pro and Ultra subscribers in Google’s Antigravity development platform.
Google says the Hydrogen update is rolling out to all Search users globally.
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.
Sarah has worked as a reporter for TechCrunch since August 2011. She joined the company after having previously spent over three years at ReadWriteWeb. Prior to her work as a reporter, Sarah worked in I.T. across a number of industries, including banking, retail and software.
You can contact or verify outreach from Sarah by emailing [email protected] or via encrypted message at sarahperez.01 on Signal.
来源:TechCrunch · AI · techcrunch.com