Workflow
Kimi K2.5
icon
Search documents
告别 AI 土味审美!Kimi K2.5 实测:扔个视频复刻 iOS 级丝滑动效
歸藏的AI工具箱· 2026-01-27 10:37
Core Insights - Kimi has launched its K2.5 model, which features enhanced aesthetic capabilities and supports multimodal recognition for videos, significantly improving the visual quality of AI-generated web pages [1][5][32] Group 1: Design Capabilities - K2.5 can better adhere to design drafts and prompts, making it easier for designers to realize their visions [8] - For non-designers, K2.5 simplifies the process by allowing users to input content without needing to find attractive design references [8] - The model has shown proficiency in replicating complex interactive components, such as a tab-switching interaction video, demonstrating its advanced multimodal and code generation capabilities [9][17] Group 2: Iterative Design Process - The iterative process with K2.5 allows for easy feedback through screenshots and annotations, leading to quick adjustments and refinements [13][19] - After several iterations, K2.5 successfully recreated a smooth animation effect for a card component system, showcasing its ability to handle multiple card types and animations [30][31] - The model can generate a design system website based on specific prompts, indicating its capability to create comprehensive design specifications [46][49] Group 3: Performance and Limitations - K2.5's performance is notably enhanced in the Agent mode, which allows for higher task completion rates by utilizing virtual machines and various tools [39] - Despite significant improvements, K2.5 still struggles with capturing precise design details, such as small corner radii and specific color values, which remains a challenge for multimodal models [66][68]
刚刚,杨植麟亲自开源Kimi K2.5!国产大模型打架的一天
机器之心· 2026-01-27 09:45
编辑 | Panda、泽南 今天真是国产大模型打架的一天!昨晚千问上新模型,今天 DeepSeek 开源 OCR 2。 中午,Kimi 也开卷,网站、App、API 开放平台和编程助手产品 Kimi Code 模型版本全面更新,Kimi K2.5 来了。 月之暗面创始人杨植麟还首次出镜,向大家分享了新模型的能力。 Kimi K2.5 是一个拥有 1 万亿参数(1 trillion)的 MoE 基础模型。相较前代,K2.5 的视觉理解能力大幅增强(可以处理视频了),Coding 能力也有了 明显提升,更重要的是,K2.5 依然开源。 Kimi K2.5 在包括 HLE、BrowseComp 和 DeepSearchQA 等极具挑战性的 agent 评测上取得了当前最佳表现(SOTA),比如 HLE(人类最后考试) 上拿到 50.2%,BrowseComp 拿到了 74.9%。 同时,K2.5 的编程能力也非常突出,它在 SWE-bench Verified 上拿到了 76.8 %,缩小了与顶尖闭源模型之间的差距,K2.5 在多项视觉理解评测上也 实现了当前开源最佳效果。 可以看到,在核心基准测试上,Kimi K ...
Kimi K2.5 上手体验:当 AI 开始学会“人海战术”,我看到了超级个体的终极形态
硬AI· 2026-01-27 09:44
杨植麟说,他们的目标是"Scale the variety of agents"。而我觉得,Kimi K2.5 最核心的价值,是 Scale your ambition(扩展你的野心)。 硬·AI 不管是GPT-5还是Claude4.5,它们确实越来越聪明,但本质上,我还是在和一个AI对话。我依然需要像个保姆一样,把任务拆碎,一步步喂给它,然后盯着 它干活。 我们都想要一个能干活的AI,但实际上更多时候它只是个知识库。 但如果,AI变成了一支随叫随到的"军队"呢? 作者 | KMGGGG、小猫 编辑 | 硬 AI 坦白说,过去半年,我对大模型的更新已经有点"审美疲劳"了。 就在刚刚,月之暗面发布了Kimi K2.5。深度体验后,我被它完全震惊到了: AI正在成为你的外包公司 。 这一次,Kimi K2.5 抛弃了单纯卷参数、卷长文本的旧叙事,直接祭出了一个让硅谷都感到压力的杀手锏:Agent Swarm(智能体集群)。 它的意义,不仅仅是评测榜单上的开源SOTA,更重要的是,它让我第一次感觉在调度一整个团队。 这,可能就是我们一直在等的"AI 2.0"时刻。 01 我的 Kimi K2.5 "指挥"体验 先说 ...