Kimi K3: Open Frontier IntelligenceKimi K3：开启前沿智能新时代

Try Kimi K3
Kimi K3 hero visual

Today, we are introducing Kimi K3 — our most capable model. Kimi K3 is a 2.8T-parameter model built on our Kimi Delta Attention and Attention Residuals, with native vision capabilities and a 1-million-token context window. It is the world's first open 3T-class model, designed for frontier intelligence across long-horizon coding, knowledge work, and reasoning.今天，我们正式推出 Kimi K3——这是我们迄今为止能力最强的模型。Kimi K3 拥有 2.8 万亿参数，基于我们自研的 Kimi Delta Attention（KDA）和 Attention Residuals（注意力残差）架构构建，具备原生视觉能力和 100 万 token 的上下文窗口。作为全球首个开源的 3T 级模型，它专为长周期编程、知识工作和复杂推理等前沿智能场景而设计。

While its overall performance still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol, Kimi K3 demonstrated frontier-level performance across our evaluation suite, consistently outperforming other tested models.尽管其整体性能相较于 Claude Fable 5 和 GPT 5.6 Sol 等最强大的闭源模型仍有差距，但 Kimi K3 在我们的评估套件中展现出了前沿水平的表现，持续超越了其他参与测试的模型。

Kimi K3 benchmark comparison
Kimi K3 benchmark comparison
Kimi K3 showcase
Kimi K3 showcase

Kimi K3 is available today on Kimi.com, Kimi Work, Kimi Code, and the Kimi API. At launch, Kimi K3 will use max thinking effort by default, with low- and high-effort modes to be introduced in subsequent updates. We are currently working closely with inference partners and open-source maintainers to align technical details and ensure a reliable rollout across the ecosystem. The full model weights will be released by July 27, 2026. Further details on the architecture, training, and evaluations will be released alongside the Kimi K3 technical report.Kimi K3 即日起在 Kimi.com、Kimi Work、Kimi Code 及 Kimi API 上线。发布初期，Kimi K3 默认采用最高思考强度模式，后续更新将引入低强度和高强度模式。目前，我们正与推理合作伙伴及开源社区维护者紧密协作，对齐技术细节，确保模型在生态系统中稳健落地。模型完整权重将于 2026 年 7 月 27 日发布。关于架构、训练和评估的更多详情，将随《Kimi K3 技术报告》一同发布。

An Open 3T-Class Model一款开源的 3T 级模型

Kimi K3 is the first open model to reach 2.8 trillion parameters. It marks the latest step in Kimi's sustained push at the scaling frontier: for nine of the past twelve months, Kimi models have set the upper bound of open-model sizes.Kimi K3 是首个达到 2.8 万亿参数的开源模型。这标志着 Kimi 在扩展前沿领域的持续突破：在过去十二个月中，有九个月 Kimi 模型都刷新了开源模型规模的上限。

Kimi K3 is built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), two architectural updates designed to improve how information flows across sequence length and model depth. We have also scaled up Mixture of Experts (MoE) sparsity, effectively activating 16 out of 896 experts when paired with a Stable LatentMoE framework. Together with refined training and data recipes, these structural changes yield an approximate 2.5× improvement in overall scaling efficiency compared to Kimi K2, allowing the model to convert compute into intelligence more effectively.Kimi K3 采用了 Kimi Delta Attention (KDA) 和 Attention Residuals (AttnRes) 这两项架构更新，旨在改善信息在序列长度和模型深度上的流动。我们还提升了专家混合模型 (MoE) 的稀疏性，在 Stable LatentMoE 框架下，可有效激活 896 个专家中的 16 个。结合精炼的训练方法和数据配方，这些结构性调整使整体扩展效率较 Kimi K2 提升了约 2.5 倍，从而使模型能更有效地将算力转化为智能。

αwKDAαwStable LatentMoEαwGated MLAαwStable LatentMoEwα3×1×Block n−1Block n−2Block n−3EmbeddingRouterLinear12123NNormLinearShared ExpertRouted ExpertLinearConvL2LinearConvL2LinearConvσσLinearσKimi Delta AttentionNormLinearOutput
Kimi K3 architecture: the Stable LatentMoE and KDA modules (left), the AttnRes operation α (top right), and the Block Attention Residuals backbone (right).Kimi K3 架构：Stable LatentMoE 和 KDA 模块（左），AttnRes 操作 α（右上），以及块注意力残差主干（右）。

Coding编程能力

Kimi K3 has strong long-horizon coding performance. Operating with minimal human oversight, it can sustain long engineering sessions, navigate massive repositories, and orchestrate terminal tools.Kimi K3 具备强大的长周期编程性能。在极少的人工干预下，它能够持续进行长时间的工程开发，驾驭海量代码库并调度终端工具。

Kimi K3 also excels in tasks blending software engineering with visual reasoning — it leverages screenshots and visuals to optimize game dev, frontend, and CAD.Kimi K3 在结合软件工程与视觉推理的任务中表现出色——它能利用截图和视觉信息来优化游戏开发、前端设计和 CAD 等工作。

The case studies below show how Kimi K3's coding capability translates into open-ended software creation and scientific research.以下案例展示了 Kimi K3 的编程能力如何转化为开放式的软件创作和科学研究成果。

Kernel Optimization内核优化

We tested the models' capability to optimize GPU kernels. Each model works independently in an identical sandbox, with up to 24 hours to profile, rewrite, and benchmark four tasks spanning AttnRes, KDA, and a 512-head-dimension MLA kernel across NVIDIA H200 and GPGPU from an alternative vendor. Kimi K3 performed competitively with Fable 5 (with fallback) and substantially outperformed Opus 4.8, GPT 5.6 Sol, and GPT 5.5.我们测试了各模型优化 GPU 内核的能力。每个模型在独立的相同沙盒中工作，拥有长达 24 小时的时间来对 NVIDIA H200 和替代厂商的 GPGPU 上的 AttnRes、KDA 以及 512 头维度 MLA 内核进行分析、重写和基准测试。Kimi K3 的表现与 Fable 5（含回退机制）相当，并大幅领先于 Opus 4.8、GPT 5.6 Sol 和 GPT 5.5。

Claude Fable 5 was evaluated by a third party, and its results may include fallback behavior. Across most models, some trajectories include small, acceptable precision shortcuts that remain within our numerical tolerance. GPGPU denotes general-purpose GPUs used for computation beyond graphics rendering.Claude Fable 5 由第三方评估，其结果可能包含回退行为。在大多数模型中，部分轨迹包含微小但可接受的精度捷径，仍在我们的数值容差范围内。GPGPU 指用于图形渲染以外计算的通用 GPU。

In the late stages of Kimi K3 development, an early version of Kimi K3 handled the majority of the team's kernel optimization works.在 Kimi K3 开发的后期阶段，早期版本的 Kimi K3 承担了团队大部分的内核优化工作。

GPU Compiler DevelopmentGPU 编译器开发

We further tested whether Kimi K3 could build a GPU programming system from scratch. Kimi K3 developed MiniTriton, a compact Triton-like compiler with its own tile-level IR layer over MLIR, optimization passes, and a PTX code-generation pipeline. Across supported roofline benchmarks, MiniTriton delivers performance on par with or better than Triton and torch.compile — beating Triton on certain workloads. Beyond microbenchmarks, MiniTriton sustains end-to-end nanoGPT training with stable convergence, the loss curve closely tracking the reference with only minor divergence — validating the full pipeline on a realistic workload. These results demonstrate that Kimi K3 can build a coherent end-to-end compiler — from DSL frontend and IR passes to PTX codegen and runtime — rather than isolated kernels; its from-scratch Tensor Core path already rivals Triton’s extensively optimized stack.我们进一步测试了 Kimi K3 从零构建 GPU 编程系统的能力。Kimi K3 开发了 MiniTriton，这是一个紧凑的类 Triton 编译器，在 MLIR 之上拥有自己的切片级 IR 层、优化通道和 PTX 代码生成流水线。在支持的屋顶线（roofline）基准测试中，MiniTriton 的性能与 Triton 和 torch.compile 持平甚至更优——在某些工作负载下超越了 Triton。除微基准测试外，MiniTriton 还支撑了端到端的 nanoGPT 训练并实现稳定收敛，损失曲线与参考值紧密吻合，仅有微小偏差，验证了其在真实工作负载下的完整流水线。这些结果证明，Kimi K3 能够构建一个连贯的端到端编译器——从 DSL 前端和 IR 通道到 PTX 代码生成和运行时——而不仅仅是孤立的内核；其从零构建的 Tensor Core 路径已足以媲美 Triton 经过深度优化的技术栈。

MiniTriton CUDA-core roofline on NVIDIA L20

Game Dev and Digital Creation游戏开发与数字创作

Kimi K3 combines strong 3D reasoning, coding, and vision capabilities to turn concepts, images, and videos into fully playable interactive experiences. Kimi K3 achieves true "vision in the loop" by seamlessly iterating between code and live screenshots—instantly seeing and refining outputs.Kimi K3 结合了强大的 3D 推理、编程和视觉能力，能将概念、图像和视频转化为完全可玩的交互式体验。Kimi K3 通过代码与实时截图之间的无缝迭代，实现了真正的“视觉闭环”，能够即时观察并精修输出结果。

Chip Design芯片设计

As an early proof of concept, Kimi K3 designed a chip to serve a nano model built on its own architecture. In a single 48-hour autonomous run, K3 built, optimized, and verified the chip using open-source EDA tools on the Nangate 45nm library. Within 4 mm², the chip closes timing at 100 MHz and sustains over 8,700 tokens/s decode throughput in simulation, packing 1.46M standard cells, 0.277 MB of SRAM, and an INT4 MAC array with fused dequantization. A chip built by a model, for a model, reflects K3's long-horizon agentic capabilities.作为早期概念验证，Kimi K3 设计了一款芯片，用于运行基于其自身架构的 nano 模型。在一次 48 小时的自主运行中，K3 使用开源 EDA 工具在 Nangate 45nm 库上构建、优化并验证了该芯片。在 4 mm² 的面积内，该芯片在 100 MHz 下实现时序收敛，在仿真中保持超过 8,700 tokens/s 的解码吞吐量，集成了 146 万个标准单元、0.277 MB SRAM 以及带有融合反量化功能的 INT4 MAC 阵列。由模型构建、为模型服务的芯片，体现了 K3 强大的长周期智能体能力。

Coding for Research科研编程

Kimi K3 bridges scientific literature and executable code, autonomously implementing, validating, and analyzing complex computational research workflows.Kimi K3 架起了科学文献与可执行代码之间的桥梁，能够自主实现、验证并分析复杂的计算研究工作流。

In one case, Kimi K3 completed in about two hours what would typically require one to two weeks of work by an experienced researcher. To reproduce the I–Love–Q universal relations in computational astrophysics, it reviewed and cross-validated 20+ papers, implemented the full numerical pipeline, evaluated 300+ equations of state, identified inconsistencies in published formulas, generated 3,000+ lines of Python code, and produced an interactive HTML dashboard for exploring the results.在一个案例中，Kimi K3 在约两小时内完成了资深研究员通常需要一到两周才能完成的工作。为了复现计算天体物理学中的 I–Love–Q 通用关系，它查阅并交叉验证了 20 多篇论文，实现了完整的数值流水线，评估了 300 多个状态方程，识别出已发表公式中的不一致之处，生成了 3,000 多行 Python 代码，并制作了一个用于探索结果的交互式 HTML 仪表板。

Knowledge Work知识工作

Kimi K3 advances end-to-end knowledge work. Beyond public benchmarks, Kimi K3 (max) demonstrates consistent gains across our internal evaluations, which are derived from recurring patterns and challenges observed in real-world user-agent workflows. These consistent advantages across distinct production-oriented workflows reflect a broad improvement in Kimi K3's agentic knowledge work capabilities.Kimi K3 推动了端到端知识工作的进步。除了公共基准测试外，Kimi K3（最高强度模式）在我们基于真实用户智能体工作流中观察到的常见模式和挑战所构建的内部评估中，表现出了持续的提升。这些在不同生产导向工作流中的持续优势，反映了 Kimi K3 在智能体知识工作能力上的全面提升。

Internal Knowledge Work Bench

Research with Interactive Visualization交互式可视化研究

Below are a few examples of what Kimi K3 in Kimi Work can produce across financial consulting and scientific research:以下是 Kimi K3 在 Kimi Work 中在金融咨询和科学研究方面所能实现的部分成果：

Case 1: Interactive 42 years of AI ASIC industry research website案例 1：AI ASIC 行业 42 年研究的交互式网站

An interactive research report you can drill into: 42 years of the ASIC industry, created through 120+ rounds of recursive self-improvement. Kimi K3 transforms evidence into bespoke charts, animated diagrams, and interactive visual narratives. It pulled data via 2.8k+ web searches/fetches and 1.1k+ terminal data pulls, across 11k+ pages spanning 87 quarterly reports and 99 original PDFs.一份可深入挖掘的交互式研究报告：通过 120 多轮递归自我提升生成的 ASIC 行业 42 年研究。Kimi K3 将证据转化为定制图表、动画图示和交互式视觉叙事。它通过 2,800 多次网络搜索/抓取和 1,100 多次终端数据拉取，处理了涵盖 87 份季度报告和 99 份原始 PDF 的 11,000 多页内容。

Case 2: Fusion Industry Research案例 2：核聚变行业研究

A consulting-style industry report with interactive visualizations—including timelines, Funnel Chart, Range Bar Chart, Gantt Charts, and publication-quality slides.一份咨询风格的行业报告，配有交互式可视化图表，包括时间轴、漏斗图、范围条形图、甘特图以及出版级的演示幻灯片。

Case 3: GWTC-5 Gravitational-wave Analysis案例 3：GWTC-5 引力波分析

An analysis of 391 gravitational-wave events using 20+ concurrent subagents, producing 7 scientific visualizations, 2 tables, and a literature synthesis from 10+ papers.利用 20 多个并发子智能体对 391 个引力波事件进行分析，生成了 7 个科学可视化图表、2 个表格，并综合了 10 多篇论文的文献综述。

Kimi K3 is also particularly effective at producing infographic-style presentations, such as the fully editable heatmap and annual report shown below:Kimi K3 在制作信息图表风格的演示文稿方面也特别高效，例如下方展示的完全可编辑热力图和年度报告：

Widgets and Dashboard小组件与仪表板

In Kimi Work, we introduce two new features - Widgets and Dashboard - which make interactions with Kimi K3 more visual and persistent. Widgets let you generate interactive components directly within a chat, with connections to local data or external plugins for continuous updates. Dashboard brings the widgets you care about most into one persistent, personalized view organized around a topic, project, or goal.在 Kimi Work 中，我们引入了“小组件 (Widgets)”和“仪表板 (Dashboard)”两项新功能，使与 Kimi K3 的交互更加直观且持久。小组件允许您直接在对话中生成交互式组件，并连接本地数据或外部插件以实现持续更新。仪表板则将您最关注的小组件整合到一个持久、个性化的视图中，围绕特定主题、项目或目标进行组织。

Video Editing视频编辑

Kimi K3 excels at motion design, animation, and video editing because its native multimodal architecture understands text, images, and video within the same model.Kimi K3 在动态设计、动画和视频编辑方面表现出色，因为其原生多模态架构能够在同一模型内理解文本、图像和视频。

In one example, K3 created a 3Blue1Brown-style motion-graphics explainer of its own architecture, translating technical ideas into animated diagrams and transitions.在一个示例中，K3 制作了一个 3Blue1Brown 风格的动态图形解释视频，介绍了其自身的架构，将技术理念转化为动画图示和转场。

In another, Kimi K3 edited its own teaser video from 56 source clips, handling clip selection, motion-matched cuts, frame-accurate beat synchronization, audio processing, and multiple rounds of revision. A high-density short video like this would typically take an experienced editor one to two working days, or a beginner three to five.在另一个案例中，Kimi K3 从 56 个源片段中剪辑了自己的预告片，处理了片段选择、动态匹配剪辑、精确到帧的节拍同步、音频处理以及多轮修改。像这样高密度的短视频，通常需要经验丰富的剪辑师工作一到两天，新手则需要三到五天。

Architecture and Infrastructure架构与基础设施

Kimi K3 is built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes). KDA provides an efficient foundation for scaling attention, while AttnRes selectively retrieves representations across depth rather than accumulating them uniformly. Together, they form the architectural backbone of a model designed to scale well beyond the trillion-parameter regime.Kimi K3 基于 Kimi Delta Attention (KDA) 和 Attention Residuals (AttnRes) 构建。KDA 为扩展注意力提供了高效基础，而 AttnRes 则选择性地跨深度检索表示，而非均匀累积。它们共同构成了该模型的架构主干，旨在实现远超万亿参数规模的扩展。

Kimi K3 uses Stable LatentMoE, effectively activating 16 of 896 experts. At this level of sparsity, routing and optimization become first-order challenges. Quantile Balancing derives expert allocation directly from router-score quantiles, eliminating heuristic updates and a sensitive balancing hyperparameter, while Per-Head Muon extends Muon by optimizing attention heads independently for more adaptive learning at scale. Sigmoid Tanh Unit (SiTU) and Gated MLA improve activation control and attention selectivity respectively. Together, these advances enable stable and efficient training at the 2.8-trillion-parameter scale.Kimi K3 使用 Stable LatentMoE，有效激活 896 个专家中的 16 个。在这一稀疏度水平下，路由和优化成为首要挑战。分位数平衡（Quantile Balancing）直接根据路由器分值分位数推导专家分配，消除了启发式更新和敏感的平衡超参数；而每头 Muon（Per-Head Muon）通过独立优化注意力头，扩展了 Muon 以实现更大规模下的自适应学习。Sigmoid Tanh Unit (SiTU) 和 Gated MLA 分别改进了激活控制和注意力选择性。这些进步共同实现了 2.8 万亿参数规模下的稳定高效训练。

Kimi K3 applies quantization-aware training from the SFT stage onward, using MXFP4 weights with MXFP8 activations for broad hardware compatibility. To prevent expert imbalance from degrading throughput at large expert-parallel scales, we introduce a fully balanced expert-parallel training method with static shapes and no host synchronization on the critical path. Since inference efficiency likewise benefits from larger high-bandwidth communication domains, we recommend deploying Kimi K3 on supernode configurations with 64 or more accelerators. Finally, as KDA poses new challenges for conventional prefix caching, we have contributed a corresponding implementation to the vLLM community, to be released alongside the model. KDA with prefill cache allows us to serve Kimi K3 at a highly competitive token price despite its scale and long context.Kimi K3 从 SFT 阶段开始应用量化感知训练，使用 MXFP4 权重和 MXFP8 激活，以实现广泛的硬件兼容性。为防止专家不平衡在大规模专家并行下降低吞吐量，我们引入了一种全平衡专家并行训练方法，采用静态形状，且关键路径上无主机同步。由于推理效率同样受益于更大的高带宽通信域，我们建议在拥有 64 个或更多加速器的超级节点配置上部署 Kimi K3。最后，由于 KDA 对传统的 KV Cache 提出了新挑战，我们已向 vLLM 社区贡献了相应的实现，将随模型一同发布。带有预填充缓存的 KDA 使我们能够在保持规模和长上下文的同时，以极具竞争力的 token 价格提供 Kimi K3 服务。

More technical details will be available in our coming report.更多技术细节将在我们即将发布的报告中提供。

Availability可用性

  • Kimi K3 Agents: Download or update to the latest Kimi app from your mobile app store, available on iOS, Android, and HarmonyOS, or visit kimi.com.Kimi K3 智能体：请从您的移动应用商店下载或更新至最新版 Kimi App（支持 iOS、Android 和鸿蒙系统），或访问 kimi.com。
  • Work with Kimi K3: Download the latest Kimi Work desktop app, version 3.1.0 or later, available for Windows and Apple silicon Macs.Kimi K3 工作台：下载最新版 Kimi Work 桌面应用（3.1.0 或更高版本），支持 Windows 和 Apple Silicon Mac。
  • Code with Kimi K3: Run Kimi Code in your terminal and select Kimi K3 using the /model command.Kimi K3 编程：在终端运行 Kimi Code，并使用 /model 命令选择 Kimi K3。
  • Build with the Kimi API: Visit the Kimi API Platform and select kimi-k3. Pricing is $0.30/MTok for cache-hit input, $3.00/MTok for cache-miss input, and $15.00/MTok for output. Powered by Mooncake's disaggregated inference architecture, the official Kimi API achieves a cache hit rate above 90% in coding workloads.Kimi API 开发：访问 Kimi API 平台并选择 kimi-k3。缓存命中输入价格为 $0.30/MTok，缓存未命中输入为 $3.00/MTok，输出为 $15.00/MTok。得益于 Mooncake 的解耦推理架构，官方 Kimi API 在编程工作负载中的缓存命中率超过 90%。
  • Bring Kimi to your organization: Kimi Enterprise provides enterprise-grade data privacy and member management, with complete separation between personal and organization accounts. Visit the pricing page and select “Get Kimi Enterprise” to subscribe for your team.企业版 Kimi：Kimi 企业版提供企业级数据隐私保护和成员管理，实现个人账户与企业账户的完全隔离。访问定价页面并选择“获取 Kimi 企业版”即可为您的团队订阅。

Full Benchmark Table完整基准测试表

Footnotes脚注

All Kimi K3 results reported below are obtained with the reasoning effort set to 'max', setting temperature = 1.0 and top-p = 1.0. Depending on the benchmark, each model is evaluated under one of three agentic harnesses — KimiCode, Claude Code, or Codex — as specified in the notes below.以下报告的所有 Kimi K3 结果均在推理强度设置为“最高 (max)”、温度 (temperature) = 1.0 且 top-p = 1.0 的条件下获得。根据基准测试的不同，每个模型都在以下注脚中指定的智能体框架（KimiCode、Claude Code 或 Codex）之一下进行评估。

Coding benchmarks编程基准测试

  1. DeepSWE. Kimi K3 is evaluated with the KimiCode harness. The GLM-5.2 score is taken from the GLM-5.2 release blog (https://z.ai/blog/glm-5.2); all remaining scores are from the official DeepSWE leaderboard (https://deepswe.datacurve.ai/), under which Kimi K3 attains 67.3 with the mini-SWE-agent harness. We report the DeepSWE v1.1 tasks.DeepSWE：Kimi K3 使用 KimiCode 框架进行评估。GLM-5.2 的分数取自 GLM-5.2 发布博客 (https://z.ai/blog/glm-5.2)；其余分数均来自 DeepSWE 官方排行榜 (https://deepswe.datacurve.ai/)，在该排行榜中，Kimi K3 使用 mini-SWE-agent 框架获得 67.3 分。我们报告的是 DeepSWE v1.1 任务。
  2. Terminal-Bench 2.1. Kimi K3 is evaluated with the KimiCode harness. For all other models, we report the best score across harnesses: GLM-5.2 with Claude Code (https://z.ai/blog/glm-5.2); Claude Opus 4.8 and Claude Fable 5 with Terminus 2 (https://artificialanalysis.ai/evaluations/terminalbench-v2-1); GPT 5.5 and GPT 5.6 Sol with Codex (https://openai.com/index/previewing-gpt-5-6-sol/).Terminal-Bench 2.1：Kimi K3 使用 KimiCode 框架进行评估。对于所有其他模型，我们报告各框架中的最佳分数：GLM-5.2 使用 Claude Code (https://z.ai/blog/glm-5.2)；Claude Opus 4.8 和 Claude Fable 5 使用 Terminus 2 (https://artificialanalysis.ai/evaluations/terminalbench-v2-1)；GPT 5.5 和 GPT 5.6 Sol 使用 Codex (https://openai.com/index/previewing-gpt-5-6-sol/)。
  3. Program Bench. Kimi K3 is evaluated with the KimiCode harness. The GLM-5.2 score is from https://z.ai/blog/glm-5.2; all other scores are from https://www.vals.ai/benchmarks/programbench.Program Bench：Kimi K3 使用 KimiCode 框架进行评估。GLM-5.2 的分数来自 https://z.ai/blog/glm-5.2；其他所有分数来自 https://www.vals.ai/benchmarks/programbench。
  4. SWE Marathon. Kimi K3, Claude Opus 4.8, and Claude Fable 5 are evaluated with the Claude Code harness; GPT-5.6 Sol is evaluated with the Codex harness. The GLM-5.2 score is from https://z.ai/blog/glm-5.2. Our evaluation is based on an H20-calibrated branch of the official v1.1 tasks (https://www.swe-marathon.org/): the Docker images, performance gates, and reference oracles for the GPU tasks have been recalibrated for H20, while the correctness and anti-cheat validators remain unchanged. Additionally, Claude Fable 5 hit fallbacks on 35% of the tasks in our evaluation, which may have negatively impacted its measured performance.SWE Marathon：Kimi K3、Claude Opus 4.8 和 Claude Fable 5 使用 Claude Code 框架进行评估；GPT-5.6 Sol 使用 Codex 框架进行评估。GLM-5.2 的分数来自 https://z.ai/blog/glm-5.2。我们的评估基于官方 v1.1 任务的 H20 校准分支 (https://www.swe-marathon.org/)：GPU 任务的 Docker 镜像、性能门限和参考预言机已针对 H20 重新校准，而正确性验证器和防作弊验证器保持不变。此外，Claude Fable 5 在我们的评估中 35% 的任务触发了回退，这可能对其测得的性能产生了负面影响。
  5. FrontierSWE. Kimi K3 is evaluated with the KimiCode harness and GPT-5.6 Sol with the Codex harness; all other results are from https://www.frontierswe.com/. Dominance scores are recomputed from the raw scores using the official evaluation script and are current as of July 16, 2026.FrontierSWE：Kimi K3 使用 KimiCode 框架评估，GPT-5.6 Sol 使用 Codex 框架评估；所有其他结果来自 https://www.frontierswe.com/。优势分数是使用官方评估脚本从原始分数重新计算得出的，截至 2026 年 7 月 16 日。
  6. PostTrain Bench. Scores for GLM-5.2, GPT-5.5, and Claude Opus 4.8 are adopted from the official PostTrainBench (https://posttrainbench.com/) results. Kimi K3, Claude Fable 5, and GPT-5.6 Sol are evaluated with the official Harbor implementation at maximum reasoning effort, averaged over three runs on H20 GPU (instead of H100 in the official setting) — Kimi K3 and Claude Fable 5 with the Claude Code harness, and GPT-5.6 Sol with the Codex harness.PostTrain Bench：GLM-5.2、GPT-5.5 和 Claude Opus 4.8 的分数采用自官方 PostTrainBench (https://posttrainbench.com/) 结果。Kimi K3、Claude Fable 5 和 GPT-5.6 Sol 在官方 Harbor 实现下以最高推理强度进行评估，并在 H20 GPU 上运行三次取平均值（而非官方设置中的 H100）——Kimi K3 和 Claude Fable 5 使用 Claude Code 框架，GPT-5.6 Sol 使用 Codex 框架。
  7. MLS Bench Lite. Kimi K3 is evaluated with the KimiCode harness; GLM-5.2 and the Claude models with the Claude Code harness; GPT-5.5 and GPT-5.6 Sol with the Codex harness.MLS Bench Lite：Kimi K3 使用 KimiCode 框架评估；GLM-5.2 和 Claude 模型使用 Claude Code 框架评估；GPT-5.5 和 GPT-5.6 Sol 使用 Codex 框架评估。
  8. KCB 2.0. Kimi K3 is evaluated with both the KimiCode and Claude Code harnesses; GLM-5.2, Claude Opus 4.8, and Claude Fable 5 with the Claude Code harness; GPT-5.5 and GPT-5.6 Sol with the Codex harness. All models are evaluated at maximum reasoning effort, except GPT-5.5, which uses the "xhigh" setting. We also note that on this in-house benchmark, 10% of the tasks entered GPT-5.6 Sol's cyber guard.KCB 2.0：Kimi K3 同时使用 KimiCode 和 Claude Code 框架评估；GLM-5.2、Claude Opus 4.8 和 Claude Fable 5 使用 Claude Code 框架评估；GPT-5.5 和 GPT-5.6 Sol 使用 Codex 框架评估。所有模型均在最高推理强度下评估，GPT-5.5 除外（使用“xhigh”设置）。我们还注意到，在此内部基准测试中，10% 的任务触发了 GPT-5.6 Sol 的网络安全防护。

Productivity and agentic benchmarks生产力与智能体基准测试

  1. For OfficeQA Pro, each test case provides the agent with the entire PDF corpus, with all PDFs rendered as images and no machine-readable text available.对于 OfficeQA Pro，每个测试用例都向智能体提供完整的 PDF 语料库，所有 PDF 均渲染为图像，且无机器可读文本。
  2. OfficeQA Pro and SpreadsheetBench 2. Kimi K3, GLM-5.2, Claude Opus 4.8, and Claude Fable 5 are evaluated with the Claude Code harness; GPT 5.5 and GPT 5.6 Sol are evaluated with the Codex harness.OfficeQA Pro 和 SpreadsheetBench 2：Kimi K3、GLM-5.2、Claude Opus 4.8 和 Claude Fable 5 使用 Claude Code 框架评估；GPT 5.5 和 GPT 5.6 Sol 使用 Codex 框架评估。
  3. MCP Atlas. All models are evaluated on the 500-task public subset with a 100-turn limit, using Gemini 3.1 Pro as the judge.MCP Atlas：所有模型均在 500 个任务的公共子集上进行评估，限制 100 轮对话，并使用 Gemini 3.1 Pro 作为裁判。
  4. AutomationBench. All models are evaluated on the 600-task public subset, following the official GitHub setup in all other respects.AutomationBench：所有模型均在 600 个任务的公共子集上进行评估，其他方面均遵循 GitHub 官方设置。
  5. BrowseComp. We adopt the context-compaction strategy used in the Claude model cards, triggered at 300K tokens. When evaluated with a 1M-token context window and no context management, Kimi K3 achieves a score of 90.4. The results of Claude Fable 5, Claude Opus 4.8, GPT 5.6 Sol, and GPT 5.5 are cited from https://www.anthropic.com/news/claude-fable-5-mythos-5 and https://openai.com/index/gpt-5-6/.BrowseComp：我们采用了 Claude 模型卡中使用的上下文压缩策略，在 300K token 时触发。当在 1M token 上下文窗口且无上下文管理的情况下进行评估时，Kimi K3 获得了 90.4 分。Claude Fable 5、Claude Opus 4.8、GPT 5.6 Sol 和 GPT 5.5 的结果引自 https://www.anthropic.com/news/claude-fable-5-mythos-5 和 https://openai.com/index/gpt-5-6/。
  6. GDPval-AA v2 and AA-Briefcase scores are cited from https://artificialanalysis.ai/.GDPval-AA v2 和 AA-Briefcase 的分数引自 https://artificialanalysis.ai/。

Multimodal benchmarks多模态基准测试

  1. Except for ZeroBench, which follows the official setting and is run five times, all multimodal scores are averaged over three runs. MMMU-Pro is evaluated following the official protocol, preserving the original input order and prepending images to the text input.除遵循官方设置并运行五次的 ZeroBench 外，所有多模态分数均为三次运行的平均值。MMMU-Pro 遵循官方协议进行评估，保留原始输入顺序并将图像置于文本输入之前。
  2. PerceptionBench. PerceptionBench is an in-house benchmark that focuses on atomic visual perception capabilities.PerceptionBench：PerceptionBench 是一项内部基准测试，专注于原子级视觉感知能力。

Limitations局限性

  1. Sensitivity to thinking history. K3 was trained in the preserved thinking history mode. If the agent harness fails to pass back all the historical thinking content as required, or if an ongoing session with another model is switched over to K3, generation quality may become highly unstable. We recommend using a harness with verified compatibility, such as Kimi Code, and avoiding switching to K3 in the middle of a session.对思考历史的敏感性：K3 是在保留思考历史模式下训练的。如果智能体框架未能按要求回传所有历史思考内容，或者在会话中途从其他模型切换到 K3，生成质量可能会变得极不稳定。我们建议使用经验证兼容的框架（如 Kimi Code），并避免在会话中途切换到 K3。
  2. Excessive proactiveness. K3's training places particular emphasis on long-horizon, challenging tasks. As a result, when it encounters minor issues or ambiguous user intent during task execution, it may make unexpected decisions on the user's behalf. If your application requires the agent to operate within well-defined boundaries and refrain from excessive improvisation, please impose more explicit behavioral constraints on K3 in the system prompt or in AGENTS.md.过度主动性：K3 的训练特别强调长周期、高难度的任务。因此，当它在执行任务时遇到小问题或用户意图模糊时，可能会代表用户做出意料之外的决定。如果您的应用要求智能体在明确的边界内运行且不进行过度即兴发挥，请在系统提示词或 AGENTS.md 中对 K3 施加更明确的行为约束。
  3. Despite being a highly competitive model overall, K3 nonetheless exhibits a noticeable gap in user experience compared with Claude Fable 5 and GPT 5.6 Sol.尽管 K3 总体上是一款极具竞争力的模型，但在用户体验方面与 Claude Fable 5 和 GPT 5.6 Sol 相比仍存在明显差距。