Chinese artificial intelligence startup Moonshot AI has launched an upgraded Kimi model, which boasts stronger coding, financial research and 3D design capabilities, while deliberately limiting its ability to carry out cyberattacks.
中国人工智能初创企业月之暗面发布升级后的 Kimi 模型，该模型代码、金融研究以及 3D 设计能力大幅增强，同时刻意限制了发起网络攻击的能力。
Kimi K3.1 Pro debuted fourth on independent evaluator Artificial Analysis' Intelligence Index, scoring 55 points, up 11 from its predecessor. It would become the highest-ranked open-weight model on the index once its weights are published, the evaluator said.
Kimi K3.1 Pro 在独立评测机构 Artificial Analysis 的智能指数榜单上初登榜单位列第四，得分为 55 分，较上一代提升 11 分。该评测机构称，模型权重一经发布，它就将成为该榜单上排名最高的开放权重模型。
The first model in Moonshot's K3.1 series is available through its Kimi apps, coding tools and developer platforms. The company plans to release its full weights — the files developers need to run the model themselves — by Nov 4.
月之暗面 K3.1 系列首款模型现已可通过 Kimi 应用、代码工具以及开发者平台使用。公司计划在 11 月 4 日前发布完整权重文件，也就是开发者本地运行该模型所需文件。
The model has 2.8 trillion parameters, with about 104 billion activated for each token processed. Its one-million-token context window allows it to handle large volumes of information in a single session.
该模型总参数量为 2.8 万亿，每处理一个词元激活约 1040 亿参数。百万词元的上下文窗口，使其能够在单次对话中处理海量信息。
Moonshot said the upgrade was built through further training of Kimi K3, with improvements aimed at completing practical tasks — from building software and creating 3D assets to analyzing financial documents and carrying out extended, multistep assignments.
月之暗面表示，本次升级基于 Kimi K3 继续训练完成，优化方向聚焦落地实用任务，包括软件开发、3D 素材生成、财报分析以及长链路、多步骤复杂任务。
The company also removed cybersecurity-related data from parts of its training pipeline, including reinforcement-learning environments, while retaining general programming tasks.
该企业还从包括强化学习环境在内的部分训练流程中剔除了网络安全相关数据，但保留通用编程任务训练。
"We aim not to release open weights that would extend the absolute frontier of cyber capability," Moonshot said.
月之暗面表示：“我们无意发布能够推高网络攻击能力上限的开放权重版本。”
On its internal test of vulnerability discovery and exploitation, the K3.1 Pro scored 81.0 percent, compared with 90.5 percent for a comparable version without such safeguards. These figures were self-reported by the company.
在内部漏洞挖掘与利用测试中，K3.1 Pro 得分 81.0%；没有加入这类安全防护的对照版本得分 90.5%。以上数据由企业自行披露。
Moonshot nevertheless reported competitive general coding performance. In its evaluation using Terminal-Bench 4.0 — which tests agents' ability to carry out practical computing tasks — K3.1 Pro scored 65.7 percent, ahead of OpenAI's GPT-6 Astra at 58.2 percent and just behind Anthropic's Claude Opus 5.5 at 66.4 percent.
尽管如此，月之暗面称该模型通用代码能力具备竞争力。在 Terminal-Bench 4.0 评测（用于测试智能体执行实际计算任务的能力）中，K3.1 Pro 得分 65.7%，高于 OpenAI 的 GPT-6 Astra（58.2%），仅次于 Anthropic 的 Claude Opus 5.5（66.4%）。
The launch comes as Moonshot prepares to tap public markets. Bloomberg reported this month that the Beijing-based firm had completed a private funding round valuing it at about $50 billion and was targeting a Hong Kong debut in the first quarter of 2027, citing people familiar with the matter.
本次模型发布正值月之暗面筹备上市之际。彭博社本月援引知情人士消息称，这家北京企业已完成一轮私募融资，估值约 500 亿美元，计划 2027 年第一季度在香港上市。
According to a report from Bloomberg Intelligence, China's top AI models have narrowed the performance gap with their United States' rivals to around 3 percent, the lowest on record.
彭博行业研究的一份报告显示，国内顶尖大模型与美国竞品之间的性能差距已经缩小至约 3%，为有记录以来最小。
Bloomberg Intelligence senior analyst Robert Lea said that the speed of the catching-up process raises fresh questions over how long the US can maintain its technological lead, and could help Chinese AI firms capture a larger share of the global market.
彭博行业研究高级分析师罗伯特・利表示，追赶速度之快，让外界重新审视美国的技术领先优势还能维持多久，同时有望助力中国 AI 企业抢占更大全球市场份额。
Stanford's 2026 AI Index said the performance gap between the top US and Chinese models had "effectively closed". As of March, the best US model was still ahead, but only by 2.7 percent, and models from the two countries had traded places at the top of rankings several times since early 2025.
斯坦福大学《2026 人工智能指数报告》指出，中美顶尖大模型之间的性能差距 “实质上已经抹平”。截至今年 3 月，美国最优模型仍小幅领先，但差距仅 2.7%；自 2025 年初以来，两国模型已经多次交替登上榜单首位。
Zhang Yaqin, founding dean of Tsinghua University's Institute for AI Industry Research, said that Chinese models have entered the global top tier, while China is increasingly carving out its own path in areas such as model efficiency, open-source development and industrial applications.
清华大学智能产业研究院创始院长张亚勤表示，中国大模型已经跻身全球第一梯队，同时中国正在模型效率、开源开发、产业落地等领域走出特色路线。
Tang Jie, a Tsinghua University computer science professor, said China still needs to catch up in high-end computing chips, fundamental AI architecture and the broader technology ecosystem, while producing more breakthroughs that originate in China rather than improving on existing approaches.
清华大学计算机科学与技术系教授唐杰表示，中国在高端算力芯片、人工智能基础架构、整体技术生态上仍有追赶空间，需要产出更多原生原创突破，而非在已有方案上迭代改良。
parameter /pəˈræmɪtə(r)/n. 参数
cyberattack /ˈsaɪbərətæk/n. 网络攻击
open-source /ˌəʊpən ˈsɔːs/adj. 开源的
architecture /ˈɑːkɪtektʃə(r)/n. 架构
