连环推 ×5推出 Kimi K3：开放前沿智能
🔹 2.8 万亿参数，100 万上下文，原生多模态
🔹 Kimi Delta Attention 在百万 token 场景下实现最高 6.3 倍解码加速
🔹 注意力残差带来约 25% 训练效率提升，额外成本低于 2%
🔹 专为长周期 Agentic 编程和自适应工作流打造
Kimi K3 现已登陆：
Kimi.com
、Kimi Work、Kimi Code 和 Kimi API
开放权重预计 2026 年 7 月 27 日
🔗 API：platform.kimi.ai
🔗 技术博客：kimi.com/blog/kimi-k3
K3 是基于 Kimi Delta Attention（KDA）和 Attention Residuals（AttnRes）打造的，这两个架构上的新设计主要为了优化信息在序列长度和模型深度上的流动方式。
同时，我们也放大了 MoE（混合专家）的稀疏度，搭配 Stable LatentMoE 框架后，每轮激活的专家数从原本的几百个变成了 896 个中的 16 个。
结合优化后的训练方案和数据配方，这些结构上的改进让整体缩放效率相比 K2 提升了大约 2.5 倍，让模型能把算力更高效地转化为智能。
内部知识工作台
除了公开基准外，Kimi K3 Max 在我们内部测试中也持续展现出提升。这些内部基准都基于真实用户代理工作流中的常见模式和挑战构建。
它在线实验基准得75.5分，DECK-Bench得73.5分，金融基准得62.6分，全面超越Claude Opus 4.8 (max)和GPT-5.5 (xhigh)。
这些成绩反映出Kimi K3在代理式知识工作能力上有明显进步，在真实场景中能提供更强大、更可靠的性能。
自我进化：AttnRes 内核优化
基于 FLA Triton AttnRes 的生产级规模（96层、8192维模型、8192 token），目标是在不改变数值结果的前提下最大化训练速度。
经过连续15小时的迭代，K3 设计出全新的两阶段内核算法，在保持数值精度的同时融合了内核参数，将正向+反向传播时间从 283.6 ms 压缩至 114.4 ms。
K3 与 Fable-5（支持潜在回退机制）达到了相近性能，但 K3 每次迭代的优化速度更快。
Kimi K3把强3D推理、编程和视觉能力结合到一起，可以把概念、图像和视频直接变成可交互的沉浸式体验。
通过代码和实时截图之间的无缝迭代，Kimi K3实现了真正的“视觉闭环”。
查看英文原文
Introducing Kimi K3: Open Frontier Intelligence
🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal
🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts
🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost
🔹 Built for long-horizon agentic coding and self-evolving workflows
Kimi K3 is now live on on
Kimi.com
, Kimi Work, Kimi Code, and the Kimi API.
Open Weights by July 27, 2026.
🔗 API:
platform.kimi.ai
🔗 Tech blog:
kimi.com/blog/kimi-k3
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 K2, allowing the model to convert compute into intelligence more effectively.
Internal knowledge work bench
Beyond public benchmarks, Kimi K3 Max also shows consistent gains on our internal benchmarks, which are built from recurring patterns and challenges in real-world user-agent workflows.
It scores 75.5 on Online Exp Bench, 73.5 on DECK-Bench, and 62.6 on Finance-Bench, outperforming Claude Opus 4.8 (max) and GPT-5.5 (xhigh) across all three.
These results reflect broad improvements in Kimi K3's agentic knowledge work capabilities, enabling more capable and reliable performance in real-world use cases.
Self-evolving: AttnRes Kernel Optimization
Given FLA Triton AttnRes at production scale (96 layers, 8192-dim model, 8192 tokens), the goal was to maximize training-side speed without changing numerics.
Over 15 hours of nonstop iteration, K3 designed a novel two-phase kernel algorithm, fused kernels while preserving numerics, and reduced forward+backward time from 283.6 ms to 114.4 ms.
K3 and Fable-5 (with potential fallback) reached similar performance, but K3 improved faster per iteration.
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
🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal
🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts
🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost
🔹 Built for long-horizon agentic coding and self-evolving workflows
Kimi K3 is now live on on
Kimi.com
, Kimi Work, Kimi Code, and the Kimi API.
Open Weights by July 27, 2026.
🔗 API:
platform.kimi.ai
🔗 Tech blog:
kimi.com/blog/kimi-k3
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 K2, allowing the model to convert compute into intelligence more effectively.
Internal knowledge work bench
Beyond public benchmarks, Kimi K3 Max also shows consistent gains on our internal benchmarks, which are built from recurring patterns and challenges in real-world user-agent workflows.
It scores 75.5 on Online Exp Bench, 73.5 on DECK-Bench, and 62.6 on Finance-Bench, outperforming Claude Opus 4.8 (max) and GPT-5.5 (xhigh) across all three.
These results reflect broad improvements in Kimi K3's agentic knowledge work capabilities, enabling more capable and reliable performance in real-world use cases.
Self-evolving: AttnRes Kernel Optimization
Given FLA Triton AttnRes at production scale (96 layers, 8192-dim model, 8192 tokens), the goal was to maximize training-side speed without changing numerics.
Over 15 hours of nonstop iteration, K3 designed a novel two-phase kernel algorithm, fused kernels while preserving numerics, and reduced forward+backward time from 283.6 ms to 114.4 ms.
K3 and Fable-5 (with potential fallback) reached similar performance, but K3 improved faster per iteration.
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
Kimi.ai认识 Kimi K3Lisan al GaibKimi-K3 的基准测试被泄露了Lisan al Gaib来自 Moonshot 的官方博客：Chubby♨️Kimi k3在设计竞技场前端排名第一。Chubby♨️Kimi K3 可能是 DeepSeek 2.0 时刻。基准测试结果已经公布，成绩非常出众。这清楚地证明了一件事…Chubby♨️Kimi K3基准测试的成绩绝了Chubby♨️@ArtificialAnlys 的独立基准评测确认了Kimi K3的性能。此外，权重即将发布。我们有一个开源、…Lisan al GaibKimi-K3 在 AAI Index 上获得 57 分，token 效率比 Opus 4.8 高出近 2 倍Ethan Mollick提个醒：我得说在对我之前的一些学术工作做复杂统计审计时，Kimi K3 Max 搞砸了不少地方，包括滥用统计和胡…Lisan al GaibKimi-K3在Text Arena排名第10Lisan al GaibKimi-K3 在 GPU 优化方面吊打 Fable 5Lisan al Gaibmoonshot 成功了Lisan al GaibKimi-K3 就是我想象中 DeepSeek-V4 怎么缩小开源/闭源差距的样子。现在该轮到 OpenAI、A…Simon WillisonKimi K3，有点调皮还有点被动攻击性：向阳乔木赞叹 Kimi K3 的美感，每个风格都是独立生成的HTML+CSS。Lisan al GaibKimi-K3在CritPt基准上得到了23%的分数karminski-牙医kimi-k3 测试速报! 前端这效果太猛了!The Rundown AIMoonshot的Kimi K3正式上线了，这可能是今年的DeepSeek时刻。AshutoshShrivastava天啦噜，Kimi K3 的基准测试成绩太疯狂了。Bindu ReddyKimi K3 其实不是 Opus 这个级别的！在长文本多轮复杂 agentic loops 上就露馅了Min ChoiKimi K3真的改变了AI模型的玩法。现在人们在创造超越ChatGPT/Claude Fable 5的各种疯狂…Ethan MollickKimi K3 写不出好谋杀悬疑小说（其他模型也一样）。这仍然是最崎岖的前沿阵地。小互月之暗面发布 Kimi K3：全球首个 3 万亿级开放模型 Bindu ReddyKimi K3基准成绩过分 - 留点怀疑看Lisan al Gaibkimi.com/blog/kimi-k3🚨 AI News | TestingCatalog🔥MOONSHOT：Moonshot AI 的 Kimi K3 在 Frontend Code Arena 排名…Orange AIK3 写作能力超过 Fable5🚨 AI News | TestingCatalogKimi K3今天首发即登陆AI/ML API，和Claude Fable 5共享同一个API key。🚨 AI News | TestingCatalogKimi K3 的基准测试在许多领域表现强劲，性能与美国领先实验室的专有模型相当。开源权重预计将于 7 月 27…Lisan al GaibKimi-K3 的博客：mp.weixin.qq.com/s/V4xhEIy8x…Lisan al Gaib现在得等2-3周看第三方的基准测试数据和token效率表现歸藏(guizang.ai)Kimi 的多模态还是很顶的，感觉随着 Claude 越来越不重视多模态理解，Kimi 这个会很有用Lisan al GaibKimi-K3在AA-Omniscience上这个体量可能还能好一点，但目前也还行。相比K2.6还是进步不小歸藏(guizang.ai)官方的测试结果AshutoshShrivastavaKimi 团队用 K3 搞出了大动作 🔥AshutoshShrivastava看看 Kimi K3 对 Fable 5 的表现。AI 领域发展太快，很难预测谁接下来会发布什么。我自己测过 K…Lisan al GaibKimi-K3在WebDev Arena排名第一AshutoshShrivastavaKimi K3 来了 🔥歸藏(guizang.ai)下面很多评论特别搞笑：“中国人一定是造了个时间机器，从明年得美国模型蒸馏的 K3。”Orange AIKimi K3 现已上架到 Cola
