Kimi-VL 技术报告

Kimi-VL Technical Report

杨植麟 Zhilin Yang · Moonshot AI · 2025-04-10 · arXiv:2504.07491 ↗ · 被引 292

打开互动全文版（逐段中英对照 + 图/公式 + 论文问答）→

摘要 · Abstract

我们提出 Kimi-VL，一个高效的开源混合专家（MoE）视觉语言模型（VLM），提供先进的多模态推理、长上下文理解和强大的智能体能力——而其语言解码器仅激活 2.8B 参数（Kimi-VL-A3B）。Kimi-VL 在具有挑战性的领域表现出色：作为通用 VLM，Kimi-VL 在多轮智能体任务（如 OSWorld）中表现出色，与旗舰模型相当。此外，它在各种具有挑战性的视觉语言任务中展现出卓越的能力，包括大学级别的图像和视频理解、OCR、数学推理和多图像理解。在对比评估中，它有效与 GPT-4o-mini、Qwen2.5-VL-7B 和 Gemma-3-12B-IT 等前沿高效 VLM 竞争，同时在多个关键领域超越 GPT-4o。Kimi-VL 还在处理长上下文和清晰感知方面取得进展。凭借 128K 扩展上下文窗口，Kimi-VL 可以处理各种长输入，在 LongVideoBench 上取得 64.5 分，在 MMLongBench-Doc 上取得 35.1 分。其原生分辨率视觉编码器 MoonViT 进一步使其能够看到并理解超高分辨率视觉输入，在 InfoVQA 上达到 83.2 分，在 ScreenSpot-Pro 上达到 34.5 分，同时保持常见任务的计算成本较低。在 Kimi-VL 的基础上，我们引入了一个高级长思考变体：Kimi-VL-Thinking-2506。通过长思维链（CoT）监督微调（SFT）和强化学习（RL）开发，最新模型展现出强大的长时推理能力（MMMU 64.0，MMMU-Pro 46.3，MathVision 56.9，MathVista 80.1，VideoMMMU 65.2），同时获得稳健的通用能力。代码和模型可在 https://github.com/MoonshotAI/Kimi-VL 公开获取。

We present Kimi-VL, an efficient open-source Mixture-of-Experts (MoE) vision-language model (VLM) that offers advanced multimodal reasoning, long-context understanding, and strong agent capabilities - all while activating only 2.8B parameters in its language decoder (Kimi-VL-A3B). Kimi-VL demonstrates strong performance across challenging domains: as a general-purpose VLM, Kimi-VL excels in multi-turn agent tasks (e.g., OSWorld), matching flagship models. Furthermore, it exhibits remarkable capabilities across diverse challenging vision language tasks, including college-level image and video comprehension, OCR, mathematical reasoning, and multi-image understanding. In comparative evaluations, it effectively competes with cutting-edge efficient VLMs such as GPT-4o-mini, Qwen2.5-VL-7B, and Gemma-3-12B-IT, while surpassing GPT-4o in several key domains. Kimi-VL also advances in processing long contexts and perceiving clearly. With a 128K extended context window, Kimi-VL can process diverse long inputs, achieving impressive scores of 64.5 on LongVideoBench and 35.1 on MMLongBench-Doc. Its native-resolution vision encoder, MoonViT, further allows it to see and understand ultra-high-resolution visual inputs, achieving 83.2 on InfoVQA and 34.5 on ScreenSpot-Pro, while maintaining lower computational cost for common tasks. Building upon Kimi-VL, we introduce an advanced long-thinking variant: Kimi-VL-Thinking-2506. Developed through long chain-of-thought (CoT) supervised fine-tuning (SFT) and reinforcement learning (RL), the latest model exhibits strong long-horizon reasoning capabilities (64.0 on MMMU, 46.3 on MMMU-Pro, 56.9 on MathVision, 80.1 on MathVista, 65.2 on VideoMMMU) while obtaining robust general abilities. Code and models are publicly accessible at https://github.com/MoonshotAI/Kimi-VL.

核心贡献 · Key contributions

• 提出 Kimi-VL，一种高效的 MoE 视觉语言模型，仅激活 2.8B 参数，实现强大的多模态推理和智能体能力。
  Proposes Kimi-VL, an efficient MoE VLM with 2.8B activated parameters, achieving strong multimodal reasoning and agent capabilities.
• 引入 MoonViT，一种原生分辨率视觉编码器，采用 2D 旋转位置编码实现高效高分辨率处理。
  Introduces MoonViT, a native-resolution vision encoder using 2D RoPE for efficient high-resolution processing.
• 通过分阶段长上下文激活实现 128K 上下文窗口，支持长视频和文档理解。
  Achieves 128K context window via staged long-context activation, enabling long video and document understanding.
• 开发 Kimi-VL-Thinking，通过长思维链监督微调和强化学习，在 MMMU、MathVision 和 MathVista 上显著提升推理能力。
  Develops Kimi-VL-Thinking with long-CoT SFT and RL, significantly improving reasoning on MMMU, MathVision, and MathVista.
• 在 OSWorld 和 WindowsAgentArena 上展示强大的智能体性能，超越 GPT-4o。
  Demonstrates strong agent performance on OSWorld and WindowsAgentArena, surpassing GPT-4o.
• 提供开源代码和模型，促进可复现性和社区进步。
  Provides open-source code and models, promoting reproducibility and community advancement.

局限 · Limitations

• 仅 2.8B 激活参数，限制了在高度专业化或语言依赖任务上的性能。
  Limited to 2.8B activated parameters, restricting performance on highly specialized or language-dependent tasks.
• 推理能力尚未达到理论上限，尤其在复杂多步推理方面。
  Reasoning capability has not reached theoretical upper bound, especially for complex multi-step inference.
• 由于注意力层参数有限，128K 上下文窗口对极长序列仍显不足。
  128K context window insufficient for extremely long sequences due to limited attention layer parameters.
• 尽管效率高，但在某些基准（如 MathVision）上仍落后于更大模型。
  Performance on some benchmarks (e.g., MathVision) lags behind larger models despite strong efficiency.
• 长思维链推理在简单任务上可能导致过度思考，如 MathVista 上的饱和现象。
  Long-CoT reasoning may lead to overthinking on simpler tasks, as seen on MathVista saturation.

论文章节 · Sections（共 14）

• 摘要 Abstract
• 1 引言 1 Introduction
• 2.1 模型架构 2.1 Model Architecture
• 2.2 Muon 优化器 2.2 Muon Optimizer
• 2.3 预训练阶段 2.3 Pre-Training Stages
• 2.4 后训练阶段 2.4 Post-Training Stages
• 2.5 基础设施 2.5 Infrastructure
• 3.1 预训练数据 3.1 Pre-Training Data
• 3.2 指令数据 3.2 Instruction Data
• 3.3 推理数据 3.3 Reasoning Data
• 4 评估 4 Evaluation
• 4.1 与最先进模型的比较 4.1 Comparison to the State-of-the-Art Models
• 4.2 Kimi-VL 的推理扩展 4.2 A Reasoning Extension of Kimi-VL
• 5 结论、局限与未来工作 5 Conclusion, Limitation, and Future Work

阅读逐段中英对照全文 →