月之暗面发布最强模型Kimi K3.1 Pro 3万亿级参数、100万Token上下文Moonshot AI unveils its most powerful model, Kimi K3.1 Pro, featuring 3 trillion parameters and 1 million token context window.


AI free token/AI免费积分
月之暗面发布最强模型Kimi K3.1 Pro 3万亿级参数、100万Token上下文

10月24日消息,月之暗面今天凌晨正式发布Kimi K3.1 Pro模型,官方称这是公司迄今能力最强的3万亿级参数模型,兼容100万Token上下文。

作为K3的后训练版本,后训练算力大幅增加,借助覆盖数百个下游领域的大规模强化学习在各领域取得全面提高,同步推出智能水平相当、输出速度约为标准版2到3倍的高速版K3.1 Pro Highspeed。

两款模型即日起借助kimi.com、Kimi App、Kimi Work、Kimi Code与Kimi API给出,完整模型权重将于11月4日前发布。

代码与长程Agent是本次升级重点,官方称在代码、Agent与知识类基准上达到或接近最强闭源模型水平,在考察真实终端受限任务的Terminal-Bench 4.0中得分65.7%,超过GPT-6 Astra的58.2%,与Claude Opus 5.5的66.4%仅有微弱差距,内部Kimi Code Bench 3.0 Long-horizon得分61.2%、高于K3的55.95%。

模型可熟练调用Unity、Unreal Engine、Godot等游戏引擎,并直接编写运行Blender、CAD脚本完成参数化3D建模,BenchCAD借助工具得分92.7%,接近Claude Opus 5.5的96.2%与GPT-6 Astra的95.9%。

文章配图

金融方向模型可端到端交付财务建模、投资分析与行业研报工作流,在外部基准OfficeQA Pro v2取得64.4的官方称当前最优成绩。

针对事实准确性的内部压力测试Kimi Hallu Bench Hard中,单项事实陈述错误率由K3的23.4%降至8.3%,含至少一处事实错误的回答比例仅为前代约42.4%,官方强调该测试刻意聚焦易错情形,错误率不代表日常使用总体水平。

安全维度,官方称在中期训练与强化学习两个阶段识别并移除网络攻防相关数据以降低滥用风险,在自研xdev-bench-hard上,未过滤版本pass@3可应对90.5%难题,最终发布版为81.0%。

本文来源:https://www.itbear.com.cn/html/2026-10/1577799.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

Moonshot AI unveils its most powerful model, Kimi K3.1 Pro, featuring 3 trillion parameters and 1 million token context window.

According to news on October 24, Moonshot AI officially released the Kimi K3.1 Pro model early this morning. The company stated that this is its most capable 3 trillion parameter model to date, supporting a 1 million token context.

As a post-training version of K3, it features significantly increased computing power for post-training, leveraging large-scale reinforcement learning that covers hundreds of downstream domains to achieve comprehensive improvements across all fields. The K3.1 Pro Highspeed version is simultaneously launched, offering comparable intelligence levels with output speed approximately 2 to 3 times faster than the standard version.

Both models are available starting today via kimi.com, the Kimi App, Kimi Work, Kimi Code, and the Kimi API. Full model weights will be released by November 4.

Code capabilities and long-horizon Agents are the focus of this upgrade. According to official claims, the model reaches or approaches the level of the strongest closed-source models on code, Agent, and knowledge benchmarks. On Terminal-Bench 4.0, which tests real terminal-constrained tasks, it scored 65.7%, surpassing GPT-6 Astra's 58.2% and trailing Claude Opus 5.5's 66.4% by only a slim margin. On the internal Kimi Code Bench 3.0 Long-horizon, it scored 61.2%, higher than K3's 55.95%.

The model can proficiently work with game engines such as Unity, Unreal Engine, and Godot, and directly write and run Blender and CAD scripts to carry out parametric 3D modeling. It scores 92.7% on BenchCAD's tool-use benchmark, close to Claude Opus 5.5's 96.2% and GPT-6 Astra's 95.9%.

文章配图

The finance-oriented model delivers financial modeling, investment analysis, and industry research workflows end-to-end, achieving a score of 64.4 on the external OfficeQA Pro v2 benchmark, which is officially recognized as the current state-of-the-art.

In an internal stress test for factual accuracy, Kimi Hallu Bench Hard, the error rate for individual factual statements dropped from 23.4% for K3 to 8.3%, and the proportion of answers containing at least one factual error was only about 42.4% of the previous generation. The company emphasized that this test deliberately focuses on error-prone scenarios, and the error rate does not represent the overall level of performance in everyday use.

In terms of safety, officials state that cybersecurity-related data was identified and removed during both the mid-training and reinforcement learning phases to reduce the risk of misuse. On their in-house developed xdev-bench-hard benchmark, the unfiltered version achieved a pass@3 rate of 90.5% on difficult problems, while the final released version achieved 81.0%.

Source: https://www.itbear.com.cn/html/2026-10/1577799.html If you have any copyright concerns, please contact us at 13951431913@139.com or submit a comment on the right.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry