跳到正文
月之暗面 Kimi 模型发布·· 2026-04-30精选AI 评分85

月之暗面开源 Kimi K2.5 原生多模态智能体模型

moonshotai/Kimi-K2.5

AI 导读

月之暗面发布开源原生多模态智能体模型 Kimi K2.5,在 Kimi-K2-Base 上以约 15 万亿视觉与文本混合 token 持续预训练而成。模型采用 MoE 架构,总参数 1T、激活参数 32B,上下文长度 256K,配备 400M 参数的 MoonViT 视觉编码器,支持即时与思考两种模式。

推荐理由

月之暗面开源原生多模态智能体模型,1T 总参数与 Agent Swarm 设计给出了与闭源旗舰对比的完整基准。

正文 · AI 翻译

Kimi K2.5


📰  技术博客     |     📄  论文

0. 更新日志

  • 2026.1.29:
    • 默认系统提示词可能会给用户带来困惑和意外行为,因此我们将其移除。
    • token <|media_start|> 不正确;它已在聊天模板中被替换为 <|media_begin|>。

1. 模型简介

Kimi K2.5 是一个开源的原生多模态智能体模型,在 Kimi-K2-Base 之上通过约 15 万亿混合视觉与文本 token 的持续预训练构建而成。它将视觉与语言理解同先进的智能体能力、即时与思考模式,以及对话与智能体范式无缝集成。

核心特性

  • 原生多模态:K2.5 在视觉—语言 token 上预训练,在视觉知识、跨模态推理以及基于视觉输入的智能体工具使用方面表现出色。
  • 视觉编程:K2.5 可根据视觉规格(UI 设计、视频工作流)生成代码,并自主编排工具进行视觉数据处理。
  • 智能体集群:K2.5 从单智能体扩展转向自主协调的集群式执行方案。它将复杂任务分解为并行子任务,由动态实例化的领域专用智能体执行。

2. 模型概要

架构 混合专家(MoE)
总参数量 1T
激活参数量 32B
层数(含稠密层) 61
稠密层数 1
注意力隐藏维度 7168
MoE 隐藏维度(每个专家) 2048
注意力头数 64
专家数量 384
每个 token 选择的专家数 8
共享专家数量 1
词表大小 160K
上下文长度 256K
注意力机制 MLA
激活函数 SwiGLU
视觉编码器 MoonViT
视觉编码器参数量 400M

3. 评测结果

基准测试 Kimi K2.5
(思考)
GPT-5.2
(xhigh)
Claude 4.5 Opus
(扩展思考)
Gemini 3 Pro
(高思考级别)
DeepSeek V3.2
(思考)
Qwen3-VL-
235B-A22B-
Thinking
推理与知识
HLE-Full 30.1 34.5 30.8 37.5 25.1† -
HLE-Full
(带工具)
50.2 45.5 43.2 45.8 40.8† -
AIME 2025 96.1 100 92.8 95.0 93.1 -
HMMT 2025(2 月) 95.4 99.4 92.9* 97.3* 92.5 -
IMO-AnswerBench 81.8 86.3 78.5* 83.1* 78.3 -
GPQA-Diamond 87.6 92.4 87.0 91.9 82.4 -
MMLU-Pro 87.1 86.7* 89.3* 90.1 85.0 -
图像与视频
MMMU-Pro 78.5 79.5* 74.0 81.0 - 69.3
CharXiv(RQ) 77.5 82.1 67.2* 81.4 - 66.1
MathVision 84.2 83.0 77.1* 86.1* - 74.6
MathVista(mini) 90.1 82.8* 80.2* 89.8* - 85.8
ZeroBench 9 9* 3* 8* - 4*
ZeroBench
(带工具)
11 7* 9* 12* - 3*
OCRBench 92.3 80.7* 86.5* 90.3* - 87.5
OmniDocBench 1.5 88.8 85.7 87.7* 88.5 - 82.0*
InfoVQA(val) 92.6 84* 76.9* 57.2* - 89.5
SimpleVQA 71.2 55.8* 69.7* 69.7* - 56.8*
WorldVQA 46.3 28.0 36.8 47.4 - 23.5
VideoMMMU 86.6 85.9 84.4* 87.6 - 80.0
MMVU 80.4 80.8* 77.3 77.5 - 71.1
MotionBench 70.4 64.8 60.3 70.3 - -
VideoMME 87.4 86.0* - 88.4* - 79.0
LongVideoBench 79.8 76.5* 67.2* 77.7* - 65.6*
LVBench 75.9 - - 73.5* - 63.6
编程
SWE-Bench Verified 76.8 80.0 80.9 76.2 73.1 -
SWE-Bench Pro 50.7 55.6 55.4* - - -
SWE-Bench Multilingual 73.0 72.0 77.5 65.0 70.2 -
Terminal Bench 2.0 50.8 54.0 59.3 54.2 46.4 -
PaperBench 63.5 63.7* 72.9* - 47.1 -
CyberGym 41.3 - 50.6 39.9* 17.3* -
SciCode 48.7 52.1 49.5 56.1 38.9 -
OJBench(cpp) 57.4 - 54.6* 68.5* 54.7* -
LiveCodeBench(v6) 85.0 - 82.2* 87.4* 83.3 -
长上下文
Longbench v2 61.0 54.5* 64.4* 68.2* 59.8* -
AA-LCR 70.0 72.3* 71.3* 65.3* 64.3* -
智能体搜索
BrowseComp 60.6 65.8 37.0 37.8 51.4 -
BrowseComp
(带上下文管理)
74.9 57.8 59.2 67.6 -
BrowseComp
(智能体集群)
78.4 - - - - -
WideSearch
(item-f1)
72.7 - 76.2* 57.0 32.5* -
WideSearch
(item-f1 智能体集群)
79.0 - - - - -
DeepSearchQA 77.1 71.3* 76.1* 63.2* 60.9* -
FinSearchCompT2&T3 67.8 - 66.2* 49.9 59.1* -
Seal-0 57.4 45.0 47.7* 45.5* 49.5* -

脚注

  1. General Testing Details
    • 我们报告了 Kimi K2.5 和 DeepSeek-V3.2 在启用思考模式下的结果、Claude Opus 4.5 在扩展思考模式下的结果、GPT-5.2 在 xhigh 推理强度下的结果,以及 Gemini 3 Pro 在高思考级别下的结果。对于视觉基准,我们还额外报告了 Qwen3-VL-235B-A22B-Thinking 的结果。
    • 除非另有说明,所有 Kimi K2.5 实验均在 temperature = 1.0、top-p = 0.95、上下文长度为 256k tokens 的条件下进行。
    • 没有公开可用分数的基准在与 Kimi K2.5 相同的条件下重新评估,并以星号(*)标记。
    • 由于服务稳定性问题,我们无法在所有基准上评估 GPT-5.2 xhigh。对于未测试的基准,我们将其标记为“-”。
  2. Text and Reasoning
    • HLE、AIME 2025、HMMT 2025(2月)和 GPQA-Diamond 的评估最大完成预算为 96k tokens。
    • AIME 和 HMMT 的结果为 32 次运行的平均值(avg@32);GPQA-Diamond 为 8 次运行的平均值(avg@8)。
    • 对于 HLE,我们报告全集(文本 & 图像)上的分数。Kimi K2.5 在无工具时得分为 31.5(文本)和 21.3(图像),在有工具时得分为 51.8(文本)和 39.8(图像)。DeepSeek-V3.2 的分数对应其纯文本子集(以 † 标记)。HF Mirror 访问被阻止,以防止潜在的数据泄漏。带工具的 HLE 使用简单的上下文管理:一旦上下文超过阈值,仅保留最新一轮的工具消息。
  3. Tool-Augmented / Agentic Search
    • 在带工具的 HLE 和所有智能体搜索基准中,Kimi K2.5 配备了搜索、代码解释器和网页浏览工具。
    • 除 BrowseComp(其中 K2.5 和 DeepSeek-V3.2 使用了 discard-all 策略)外,未应用上下文管理,超过支持上下文长度的任务直接计为失败。
    • 测试系统提示强调深入且主动地使用工具,指示模型仔细推理、利用工具并验证不确定的信息。完整提示将在技术报告中提供。
    • Seal-0 和 WideSearch 的结果为四次运行的平均值(avg@4)。
  4. Vision Benchmarks
    • Max-tokens = 64k,为三次运行的平均值(avg@3)。
    • ZeroBench(带工具)使用 max-tokens-per-step = 24k 和 max-steps = 30 进行多步推理。
    • MMMU-Pro 遵循官方协议,保留输入顺序并在前面添加图像。
    • GPT-5.2-xhigh 约有 10% 的失败率(重试 3 次后仍无输出),视为错误;报告的分数可能低估了真实性能。
    • WorldVQA,一个旨在评估原子级视觉中心世界知识的基准。访问 WorldVQA:https://github.com/MoonshotAI/WorldVQA。
    • OmniDocBench 分数计算为 (1 − 归一化 Levenshtein 距离) × 100,分数越高表示准确性越优。
  5. Coding Tasks
    • Terminal-Bench 2.0 分数使用默认智能体框架(Terminus-2)和提供的 JSON 解析器获得。在我们的实现中,我们在非思考模式下评估了 Terminal-Bench 2.0。做出这一选择是因为我们当前针对思考模式的上下文管理策略与 Terminus-2 不兼容。
    • 对于 SWE-Bench 系列评估(包括 verified、multilingual 和 pro),我们使用了内部开发的评估框架。该框架包含一组最小工具——bash 工具、createfile 工具、insert 工具、view 工具、strreplace 工具和 submit 工具——以及为任务量身定制的系统提示。最高分在非思考模式下取得。
    • Claude Opus 4.5 在 CyberGym 上的分数是在非思考设置下报告的。
    • 所有报告的编码任务分数均为 5 次独立运行的平均值。
  6. Long-Context Benchmarks
    • AA-LCR:分数为三次运行的平均值(avg@3)。
    • LongBench-V2:相同的提示词和输入上下文统一标准化为约 128k tokens。
  7. Agent Swarm
    • BrowseComp(Swarm 模式):主智能体最多 15 步;子智能体最多 100 步。
    • WideSearch(Swarm 模式):主智能体和子智能体最多 100 步。

4. 原生 INT4 量化

Kimi-K2.5 采用与 Kimi-K2-Thinking 相同的原生 int4 量化方法。

5. 部署

你可以在 https://platform.moonshot.ai 上访问 Kimi-K2.5 的 API,我们为你提供兼容 OpenAI/Anthropic 的 API。为验证部署是否正确,我们还提供了 Kimi Vendor Verifier。 目前,建议在以下推理引擎上运行 Kimi-K2.5:

  • vLLM
  • SGLang
  • KTransformers

transformers 的最低版本要求为 4.57.1。

部署示例可在 模型部署指南 中找到。


6. 模型使用

以下使用示例演示如何调用我们的官方 API。

对于使用 vLLM 或 SGLang 部署的第三方 API,请注意:

  • 视频内容对话是一项实验性功能,目前仅在我们的官方 API 中支持。

  • 推荐的 temperature 在 Thinking 模式下为 1.0,在 Instant 模式下为 0.6。

  • 推荐的 top_p 为 0.95。

  • 要使用 instant 模式,你需要在 extra_body 中传入 {'chat_template_kwargs': {"thinking": False}}。

对话补全

这是一个简单的对话补全脚本,展示如何在 Thinking 和 Instant 模式下调用 K2.5 API。

import openai
import base64
import requests
def simple_chat(client: openai.OpenAI, model_name: str):
    messages = [
        {'role': 'system', 'content': 'You are Kimi, an AI assistant created by Moonshot AI.'},
        {
            'role': 'user',
            'content': [
                {'type': 'text', 'text': 'which one is bigger, 9.11 or 9.9? think carefully.'}
            ],
        },
    ]
    response = client.chat.completions.create(
        model=model_name, messages=messages, stream=False, max_tokens=4096
    )
    print('====== Below is reasoning_content in Thinking Mode ======')
    print(f'reasoning content: {response.choices[0].message.reasoning_content}')
    print('====== Below is response in Thinking Mode ======')
    print(f'response: {response.choices[0].message.content}')

    # To use instant mode, pass {"thinking" = {"type":"disabled"}}
    response = client.chat.completions.create(
        model=model_name,
        messages=messages,
        stream=False,
        max_tokens=4096,
        extra_body={'thinking': {'type': 'disabled'}},  # this is for official API
        # extra_body= {'chat_template_kwargs': {"thinking": False}}  # this is for vLLM/SGLang
    )
    print('====== Below is response in Instant Mode ======')
    print(f'response: {response.choices[0].message.content}')

带视觉内容的对话补全

K2.5 支持图像和视频输入。

以下示例演示如何以图像输入调用 K2.5 API:

import openai
import base64
import requests

def chat_with_image(client: openai.OpenAI, model_name: str):
    url = 'https://hf-mirror.com/moonshotai/Kimi-K2.5/resolve/main/figures/kimi-logo.png'
    image_base64 = base64.b64encode(requests.get(url).content).decode()
    messages = [
        {
            'role': 'user',
            'content': [
                {
                    'type': 'image_url',
                    'image_url': {'url': f'data:image/png;base64, {image_base64}'},
                },
                {'type': 'text', 'text': 'Describe this image in detail.'},
            ],
        }
    ]

    response = client.chat.completions.create(
        model=model_name, messages=messages, stream=False, max_tokens=8192
    )
    print('====== Below is reasoning_content in Thinking Mode ======')
    print(f'reasoning content: {response.choices[0].message.reasoning_content}')
    print('====== Below is response in Thinking Mode ======')
    print(f'response: {response.choices[0].message.content}')

    # Also support instant mode if you pass {"thinking" = {"type":"disabled"}}
    response = client.chat.completions.create(
        model=model_name,
        messages=messages,
        stream=False,
        max_tokens=4096,
        extra_body={'thinking': {'type': 'disabled'}},  # this is for official API
        # extra_body= {'chat_template_kwargs': {"thinking": False}}  # this is for vLLM/SGLang
    )
    print('====== Below is response in Instant Mode ======')
    print(f'response: {response.choices[0].message.content}')

    return response.choices[0].message.content

以下示例演示如何以视频输入调用 K2.5 API:

import openai
import base64
import requests

def chat_with_video(client: openai.OpenAI, model_name:str):
    url = 'https://hf-mirror.com/moonshotai/Kimi-K2.5/resolve/main/figures/demo_video.mp4'
    video_base64 = base64.b64encode(requests.get(url).content).decode()
    messages = [
        {
            "role": "user",
            "content": [
                {
                    "type": "video_url",
                    "video_url": {"url": f"data:video/mp4;base64,{video_base64}"},
                },
                {"type": "text","text": "Describe the video in detail."},
            ],
        }
    ]

    response = client.chat.completions.create(model=model_name, messages=messages)
    print('====== Below is reasoning_content in Thinking Mode ======')
    print(f'reasoning content: {response.choices[0].message.reasoning_content}')
    print('====== Below is response in Thinking Mode ======')
    print(f'response: {response.choices[0].message.content}')

    # Also support instant mode if pass {"thinking" = {"type":"disabled"}}
    response = client.chat.completions.create(
        model=model_name,
        messages=messages,
        stream=False,
        max_tokens=4096,
        extra_body={'thinking': {'type': 'disabled'}},  # this is for official API
        # extra_body= {'chat_template_kwargs': {"thinking": False}}  # this is for vLLM/SGLang
    )
    print('====== Below is response in Instant Mode ======')
    print(f'response: {response.choices[0].message.content}')
    return response.choices[0].message.content

交错思考与多步工具调用

K2.5 与 K2 Thinking 采用相同的交错思考与多步工具调用设计。使用示例请参阅 K2 Thinking 文档。

编码智能体框架

Kimi K2.5 与 Kimi Code CLI 作为其智能体框架配合使用时效果最佳——欢迎在 https://www.kimi.com/code 试用。


7. 许可证

代码仓库和模型权重均依据 Modified MIT License 发布。


8. 第三方声明

参见 THIRD PARTY NOTICES


9. 联系我们

如有任何问题,请通过 [email protected] 联系我们。

10. 参考文献

如果你认为 K2.5 对你的研究有帮助,请按如下方式引用 K2.5 技术报告:

@misc{kimiteam2026kimik25visualagentic,
      title={Kimi K2.5: Visual Agentic Intelligence}, 
      author={Kimi Team and Tongtong Bai and Yifan Bai and Yiping Bao and S. H. Cai and Yuan Cao and Y. Charles and H. S. Che and Cheng Chen and Guanduo Chen and Huarong Chen and Jia Chen and Jiahao Chen and Jianlong Chen and Jun Chen and Kefan Chen and Liang Chen and Ruijue Chen and Xinhao Chen and Yanru Chen and Yanxu Chen and Yicun Chen and Yimin Chen and Yingjiang Chen and Yuankun Chen and Yujie Chen and Yutian Chen and Zhirong Chen and Ziwei Chen and Dazhi Cheng and Minghan Chu and Jialei Cui and Jiaqi Deng and Muxi Diao and Hao Ding and Mengfan Dong and Mengnan Dong and Yuxin Dong and Yuhao Dong and Angang Du and Chenzhuang Du and Dikang Du and Lingxiao Du and Yulun Du and Yu Fan and Shengjun Fang and Qiulin Feng and Yichen Feng and Garimugai Fu and Kelin Fu and Hongcheng Gao and Tong Gao and Yuyao Ge and Shangyi Geng and Chengyang Gong and Xiaochen Gong and Zhuoma Gongque and Qizheng Gu and Xinran Gu and Yicheng Gu and Longyu Guan and Yuanying Guo and Xiaoru Hao and Weiran He and Wenyang He and Yunjia He and Chao Hong and Hao Hu and Jiaxi Hu and Yangyang Hu and Zhenxing Hu and Ke Huang and Ruiyuan Huang and Weixiao Huang and Zhiqi Huang and Tao Jiang and Zhejun Jiang and Xinyi Jin and Yu Jing and Guokun Lai and Aidi Li and C. Li and Cheng Li and Fang Li and Guanghe Li and Guanyu Li and Haitao Li and Haoyang Li and Jia Li and Jingwei Li and Junxiong Li and Lincan Li and Mo Li and Weihong Li and Wentao Li and Xinhang Li and Xinhao Li and Yang Li and Yanhao Li and Yiwei Li and Yuxiao Li and Zhaowei Li and Zheming Li and Weilong Liao and Jiawei Lin and Xiaohan Lin and Zhishan Lin and Zichao Lin and Cheng Liu and Chenyu Liu and Hongzhang Liu and Liang Liu and Shaowei Liu and Shudong Liu and Shuran Liu and Tianwei Liu and Tianyu Liu and Weizhou Liu and Xiangyan Liu and Yangyang Liu and Yanming Liu and Yibo Liu and Yuanxin Liu and Yue Liu and Zhengying Liu and Zhongnuo Liu and Enzhe Lu and Haoyu Lu and Zhiyuan Lu and Junyu Luo and Tongxu Luo and Yashuo Luo and Long Ma and Yingwei Ma and Shaoguang Mao and Yuan Mei and Xin Men and Fanqing Meng and Zhiyong Meng and Yibo Miao and Minqing Ni and Kun Ouyang and Siyuan Pan and Bo Pang and Yuchao Qian and Ruoyu Qin and Zeyu Qin and Jiezhong Qiu and Bowen Qu and Zeyu Shang and Youbo Shao and Tianxiao Shen and Zhennan Shen and Juanfeng Shi and Lidong Shi and Shengyuan Shi and Feifan Song and Pengwei Song and Tianhui Song and Xiaoxi Song and Hongjin Su and Jianlin Su and Zhaochen Su and Lin Sui and Jinsong Sun and Junyao Sun and Tongyu Sun and Flood Sung and Yunpeng Tai and Chuning Tang and Heyi Tang and Xiaojuan Tang and Zhengyang Tang and Jiawen Tao and Shiyuan Teng and Chaoran Tian and Pengfei Tian and Ao Wang and Bowen Wang and Chensi Wang and Chuang Wang and Congcong Wang and Dingkun Wang and Dinglu Wang and Dongliang Wang and Feng Wang and Hailong Wang and Haiming Wang and Hengzhi Wang and Huaqing Wang and Hui Wang and Jiahao Wang and Jinhong Wang and Jiuzheng Wang and Kaixin Wang and Linian Wang and Qibin Wang and Shengjie Wang and Shuyi Wang and Si Wang and Wei Wang and Xiaochen Wang and Xinyuan Wang and Yao Wang and Yejie Wang and Yipu Wang and Yiqin Wang and Yucheng Wang and Yuzhi Wang and Zhaoji Wang and Zhaowei Wang and Zhengtao Wang and Zhexu Wang and Zihan Wang and Zizhe Wang and Chu Wei and Ming Wei and Chuan Wen and Zichen Wen and Chengjie Wu and Haoning Wu and Junyan Wu and Rucong Wu and Wenhao Wu and Yuefeng Wu and Yuhao Wu and Yuxin Wu and Zijian Wu and Chenjun Xiao and Jin Xie and Xiaotong Xie and Yuchong Xie and Yifei Xin and Bowei Xing and Boyu Xu and Jianfan Xu and Jing Xu and Jinjing Xu and L. H. Xu and Lin Xu and Suting Xu and Weixin Xu and Xinbo Xu and Xinran Xu and Yangchuan Xu and Yichang Xu and Yuemeng Xu and Zelai Xu and Ziyao Xu and Junjie Yan and Yuzi Yan and Guangyao Yang and Hao Yang and Junwei Yang and Kai Yang and Ningyuan Yang and Ruihan Yang and Xiaofei Yang and Xinlong Yang and Ying Yang and Yi Yang and Yi Yang and Zhen Yang and Zhilin Yang and Zonghan Yang and Haotian Yao and Dan Ye and Wenjie Ye and Zhuorui Ye and Bohong Yin and Chengzhen Yu and Longhui Yu and Tao Yu and Tianxiang Yu and Enming Yuan and Mengjie Yuan and Xiaokun Yuan and Yang Yue and Weihao Zeng and Dunyuan Zha and Haobing Zhan and Dehao Zhang and Hao Zhang and Jin Zhang and Puqi Zhang and Qiao Zhang and Rui Zhang and Xiaobin Zhang and Y. Zhang and Yadong Zhang and Yangkun Zhang and Yichi Zhang and Yizhi Zhang and Yongting Zhang and Yu Zhang and Yushun Zhang and Yutao Zhang and Yutong Zhang and Zheng Zhang and Chenguang Zhao and Feifan Zhao and Jinxiang Zhao and Shuai Zhao and Xiangyu Zhao and Yikai Zhao and Zijia Zhao and Huabin Zheng and Ruihan Zheng and Shaojie Zheng and Tengyang Zheng and Junfeng Zhong and Longguang Zhong and Weiming Zhong and M. Zhou and Runjie Zhou and Xinyu Zhou and Zaida Zhou and Jinguo Zhu and Liya Zhu and Xinhao Zhu and Yuxuan Zhu and Zhen Zhu and Jingze Zhuang and Weiyu Zhuang and Ying Zou and Xinxing Zu},
      year={2026},
      eprint={2602.02276},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2602.02276}, 
}

来源:月之暗面 Kimi 模型发布 · hf-mirror.com