谷歌旗下 DeepMind 团队正式宣布推出最新前沿大模型 Gemini 4 Carbon 1。该模型在 Artificial Analysis Intelligence Index v5.0 评测中位列第一，得分为 59 1。Gemini 4 Carbon 的核心优势体现在网络安全和长周期任务处理上，其在关键安全基准测试中表现显著，具体得分包括 SEC-bench Pro 的 88.2%、ExploitGym 的 54.8% 以及 CWE-bench 的 71.0% 1。在 Web 开发领域，该模型在 WebDev Arena 排名第 4，Elo 积分为 1769，较之前提升了 89 分 1。
在安全性方面，Gemini 4 Carbon 展现出极强的抗攻击能力，其 Terminal-Bench 4.0 作弊成功率及间接提示注入攻击成功率均降至 0.0% 1。内部对比数据显示，该模型生成安全代码的成功率比 Argon 提升超过 2 倍 1。此外，Chrome 团队报告称其补丁质量优于其他实验室模型 10 倍以上 1。目前，该模型正通过公平风计划（Fairwind Program）向部分网络防御者提供访问权限 1。谷歌计划以每百万输入 Token 2 美元、输出 Token 10 美元的初始价格，逐步面向开发者及消费者推出该模型 1。
Google DeepMind has announced the launch of its latest frontier large language model, Gemini 4 Carbon 1. The model currently holds the number one position on the Artificial Analysis Intelligence Index v5.0 with a score of 59 1.
Gemini 4 Carbon demonstrates significant advancements in cybersecurity and long-horizon task capabilities 1. It achieved scores of 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench 1. Additionally, the model ranks fourth on the WebDev Arena benchmark with an Elo rating of 1769, marking an increase of 89 points 1.
In terms of safety and reliability, Gemini 4 Carbon reduces the success rate of cheating on Terminal-Bench 4.0 and indirect prompt injection attacks to 0.0% 1. Internal comparisons reveal that the model generates secure code with a success rate more than twice that of Argon, while the Chrome team reported patch quality exceeding that of other laboratory models by over ten times 1.
Currently, access is being provided to select cyber defenders through the Fairwind Program 1. The model plans to roll out to developers and consumers at an initial price of $2 per million input tokens and $10 per million output tokens 1.
评论
还没有评论，欢迎留下第一条。