Google 发布 Gemini 4 Hydrogen,称在独立智能指数上超过 Argon
Google releases Gemini 4 Hydrogen, a lower-cost model said to outscore Argon - SiliconANGLE
Google 发布 Gemini 4 家族中成本更低的 Gemini 4 Hydrogen,称其在 Artificial Analysis Intelligence Index 默认高思考档位下得 51 分,高于 Gemini 4 Argon 的 50 分,新增的 xhigh 档位可到 54 分。该模型跳过了 Argon 面向 Fairwind Program 安全人员的近两周先行阶段,即日起向所有用户开放,包括 Gemini 应用免费版用户。Google 称在 Terminal-Bench 4.1 上 Hydrogen 得分 57%,Gemini 3.8 Flash 为 18.2%;同日加入 Antigravity 编码工具的 Orchestrator 会在 Hydrogen 与 Gemini 4 Carbon 之间切换,在 Terminal-Bench 4.0 上以平均少 24% 的配额接近 Carbon 的成功率。
推荐理由
精选AI 评分 84
原文给出 Hydrogen 与 Argon 的分数对比和开放节奏差异,可据此判断 Gemini 4 家族的成本与能力取舍。
主题
正文
Google LLC today released Gemini 4 Hydrogen, a lower-cost model in its Gemini 4 artificial intelligence family that it says outscores its frontier Gemini 4 Argon model on an independent intelligence ranking.
Measured on the Artificial Analysis Intelligence Index at its default high thinking level, Hydrogen scored 51 to Argon’s 50, according to Google. A fourth thinking level called xhigh, new with Hydrogen, takes the score to 54.
The rollout is different, too. Argon spent nearly two weeks with members of Google’s Fairwind Program, a group of vetted cybersecurity defenders, before paying developers could get at it on Oct. 13. Hydrogen skipped that stage. Anyone can use it from today, including people on the free version of the Gemini app.
Koray Kavukcuoglu, who became head of Google DeepMind in August when Demis Hassabis moved into a chairman role, wrote in the announcement that Hydrogen is “built for rapid execution.”
Orchestrator, which Google added to its Antigravity coding tool today, picks between Hydrogen and Gemini 4 Carbon as a session goes along. Google’s runs of the Terminal-Bench 4.0 coding test found Orchestrator getting close to Carbon’s success rate on an average of 24% less quota.
Gemini 3.8 Flash, which came out in September, is the yardstick for most of the other numbers in the announcement. Against it, Hydrogen finished tasks 2.5 times faster at the default setting. On Terminal-Bench 4.1, which tests work done at a command line, Hydrogen more than tripled the older model’s 18.2% with a score of 57%.
Among the early testers Google quoted, Andrew Carr, a former OpenAI scientist who is now chief scientist at 3D animation startup Cartwheel Inc., said the model caught two problems with an approach his team had been stuck on for 10 days. Discovery Loop Inc., the startup former Google Chief Scientist Jeff Dean co-founded after leaving in August, was another.
Image: Google