Google 于2026年11月13日正式推出 Gemini 4 系列中的新模型“Hydrogen”,该模型旨在替代广泛使用的 Gemini 3.8 Flash 14。作为基于 Gemini 4 Carbon 微调的模型,Hydrogen 侧重于日常编码和智能体(Agentic)工作流的速度与效率 12。此次发布被视为 Google 搜索的重大升级,新模型现已默认启用并处理最复杂的查询 57。在可用性方面,尽管有报道称其曾处于内部测试阶段 8,但多方信源确认 Hydrogen 即日起通过 API、Gemini App、Google Search 等渠道向全球所有层级用户开放 1457。
性能基准显示,Hydrogen 在执行速度和任务完成率上显著优于前代产品。相比 Gemini 3.8 Flash,Hydrogen 的任务完成速度提高了2.5倍 145,且在 Terminal-Bench 4.1 测试中得分从18.2%跃升至57% 134。在 Artificial Analysis Intelligence Index 上,Hydrogen 默认高级思考模式得分为51,略高于 Gemini 4 Argon 的50分;而在超高思维模式(xhigh)下得分可达54 145。此外,该模型将“虚构声明”减少了超过5倍,并将代理任务的过度行动率降低了高达81%,内部诚实评估率达到94% 14。
为了优化多模型协作,Google 在 Antigravity 平台引入了名为“Orchestrator”的新功能 13。该功能通过动态双向调度机制,结合高性能的 Gemini 4 Carbon 和轻量级的 Hydrogen,以平衡任务质量、效率和延迟 3。数据显示,这一组合平均减少24%的配额消耗,同时提升19%的执行速度,并在 Terminal-Bench 中实现了接近 Carbon 的任务成功率 13。Hydrogen 支持最高1M token的上下文窗口,知识截止日期为2026年3月 12。定价方面,输入费用为每百万 token 1.50美元,输出费用为每百万 token 7.50美元 1。安全评估指出,Hydrogen 在化学、生物、放射、核(CBRN)及网络安全领域未达到临界能力级别,但触发了警戒阈值 2。
Google released Gemini 4 Hydrogen on November 13, 2026 1. This new model officially replaces the widely used Gemini 3.8 Flash 468 and is positioned to prioritize fast execution and daily coding tasks 12. It serves as a lightweight counterpart to the more powerful Gemini 4 Carbon and Argon models 12. Unlike previous iterations which had stricter access limitations, Hydrogen was made immediately available to all users globally through platforms including Google Search, the Gemini App, API access, and AI Studio 147.
The launch represents a significant upgrade for search functionality. Robby Stein, VP of Product at Google Search, stated that Hydrogen is "now the new default model in AI Mode & being used for your hardest questions in AI Overviews" 7. The model is designed to handle complex, multi-part queries, typos, and highly specific requests by breaking them down into point-by-point answers 57. In benchmark comparisons against its predecessor, 3.8 Flash, Hydrogen demonstrated a 2.5x increase in task completion speed under high thinking modes 145. On the Terminal-Bench 4.1 test, which measures code generation capabilities, Hydrogen scored 57%, a substantial improvement from the 18.2% score achieved by Gemini 3.8 Flash 134.
Performance metrics indicate that Hydrogen rivals or exceeds higher-tier models in specific contexts. On the Artificial Analysis Intelligence Index, the model achieved a score of 51 in its default advanced thinking mode, slightly surpassing the score of 50 recorded by the stronger Gemini 4 Argon 135. When configured in its highest thinking mode (xhigh), the index score rose to 54 4. The model offers four distinct thinking levels—low, medium, high (default), and xhigh—to allow users to balance computational resources with output quality 5. Additionally, safety assessments show that Hydrogen reduces hallucinated statements by more than five times compared to 3.8 Flash, while also decreasing over-action in agentic tasks by up to 81% 14.
A key feature accompanying the release is the Orchestrator function, introduced within the Google Antigravity platform 13. This tool dynamically switches between the high-performance Gemini 4 Carbon and the efficient Hydrogen model to optimize workflow latency and cost 3. According to reports, this orchestration approach results in a 24% reduction in quota consumption and a 19% increase in execution speed while maintaining near-Carbon level success rates on terminal benchmarks 13. While one source notes that these features are currently rolling out to paid Google AI Pro and Ultra subscribers 3, other outlets confirm immediate availability across all user tiers via standard APIs and consumer apps 147.
Technical documentation reveals that Hydrogen is fine-tuned from the base Gemini 4 Carbon architecture and supports a context window of up to 1 million tokens for combined input and output 2. It handles multimodal inputs and outputs, with a default synchronous token limit of 64K, though background decoding can reach the full 1M capacity 2. The model’s knowledge cutoff date is listed as March 2026 2. Security evaluations indicate that while Hydrogen has reached warning thresholds in Chemical, Biological, Radiological, Nuclear (CBRN) and cybersecurity domains, it has not attained Critical Capability Levels (CCLs) 2. Furthermore, the model ranks fifth on Agent Arena for multi-step tasks such as deep research and data analysis 5.
评论
还没有评论,欢迎留下第一条。