Kimi

国产 AI 智能助手，超长上下文窗口和中文理解能力著称

large language model

https://www.kimi.com/

Visit Website

In-depth Report

Kimi is a domestic AI intelligent assistant developed by Beijing Dark Side of the Moon Technology Co., Ltd. and will be officially released in October 2023. In May 2026, Dark Side of the Moon completed a new round of financing of approximately US$2 billion, with a valuation exceeding US$20 billion, making it one of the most highly valued AI unicorns in the country. Kimi is famous for its ultra-long context window (2 million tokens) and Chinese understanding ability, and occupies an important position in the domestic AI market. The Kimi K2.6 programming model released in April 2026 surpassed GPT-5.4 and Claude Opus 4.6 in multiple benchmark tests. Kimi provides a free version and a professional paid version, which is suitable for users who need long document processing and Chinese and English conversations, but there is still a gap between coding capabilities and creative writing compared to ChatGPT and Claude.
Beijing Dark Side of the Moon Technology Co., Ltd. (Moonshot AI) was established in 2023 and focuses on the research and development of large language models. Although company founder Yann LeCun is not directly involved, team members come from top AI research institutions at home and abroad. The name of Dark Side of the Moon is taken from the astronomical concept of "Dark Side of the Moon", which symbolizes the spirit of exploration into unknown areas. The company has raised more than US$3.9 billion in total, with investors including Meituan Longzhu, Hillhouse Capital, Tencent Investment and other well-known institutions. As the company's core product, Kimi is positioned as an "AI assistant" and aims to provide users with services such as intelligent dialogue, long document processing, and information retrieval. In July 2025, Dark Side of the Moon released the Kimi K2 open source large model, which has 1 trillion total parameters (32 billion activation parameters) and is specially designed for "intelligent agent" capabilities. Kimi K2.5 will be released in January 2026, with 2 million tokens and ultra-long context, supporting multi-modal understanding. Kimi K2.6 will be released in April 2026 and achieve leading results in programming benchmark tests. The company's product line also includes Kimi Code (programming assistant) and Kimi Claw (Agent tool).
Kimi provides rich AI functions covering multiple scenarios in work and life. Core functions include: PPT generation (AI generated presentations), document processing (long document reading comprehension and summary), in-depth research (professional field research analysis), website generation (AI created website), form processing (data analysis and management), Agent cluster (multi-Agent collaboration system), Kimi Code (programming code assistant) and Agent (customized agent). Kimi's ability to handle long documents is one of its core competencies. Kimi K2.5 supports 2 million token ultra-long contexts and can read entire books, financial reports, papers and other long texts at once, and conduct in-depth analysis and summary. In actual use, Kimi can quickly locate key information in long documents, extract summaries, and conduct questions and answers about the document content. This feature is of high practical value for researchers, students, and professionals who need to process large amounts of literature. Multi-modal capabilities are one of the core breakthroughs of Kimi K2.5. For a long time, the training of AI models has often treated text and vision as two independent modules, which can easily lead to the separation of the model's capabilities between different modalities. Kimi K2.5 achieves true "joint optimization" of text and vision, capable of understanding image content and performing cross-modal reasoning. This ability is particularly useful when processing research reports containing charts and screenshots. Kimi's programming abilities continue to evolve. Kimi K2.6 outperforms GPT-5.4 and Claude Opus 4.6 on programming benchmarks such as SWE-Bench Pro and Terminal-Bench. Kimi can quickly read API documents, locate required information, and generate code based on user needs, assisting in rapid debugging and improving programming efficiency. For non-professional programmers, Kimi Code lowers the barriers to programming. Kimi's interface is designed to be simple and easy to use. Users can visit Kimi.ai directly or use Kimi in a search engine. The free version is open to all users, can be used without circumventing the firewall, and is very friendly to domestic users. The Pro version offers a longer context window and faster response times.
Kimi adopts a freemium business model and provides a free version and a professional paid version. The free version provides basic AI conversation and document processing functions and is suitable for light users. The Pro version offers longer context windows, faster response times, and more advanced features. In terms of API pricing, Kimi has a high price/performance ratio in the industry. According to the official pricing document, the API prices of Kimi K2 series models are as follows (per million tokens): kimi-k2-0905-preview input cache hit ¥1.00, miss ¥4.00, output ¥16.00; kimi-k2-turbo-preview input cache hit ¥1.00, miss ¥8.00, output ¥58.00 (high-speed version, output speed reaches 60-100 tokens/second); kimi-k2-thinking input cache hit ¥1.00, miss ¥4.00, output ¥16.00. Compared with similar domestic products, Kimi's pricing is competitive. Compared with competing products such as MiniMax, Kimi's free quota is more generous and suitable for developers to conduct early testing and prototype development.
Judging from user feedback in Zhihu and other communities, Kimi has a good reputation in the domestic AI assistant market. Some users commented that "Kimi is currently the best among domestic AI assistants. Its Chinese understanding and long document processing have reached the top level, and it is completely free and does not require circumventing the firewall." This evaluation reflects Kimi’s core advantages among domestic users: free availability, strong Chinese capabilities, and excellent long document processing. The advantages of Kimi are mainly reflected in the following aspects: fast response time, compared with competing products such as DeepSeek, Kimi's response time is shorter; the output is concise, not much nonsense, and high efficiency; the function is focused, although it cannot generate pictures, but it is very suitable for searchers who need to find solutions quickly in work scenarios; the long document processing ability is strong, and it can process a large amount of text at one time and extract key information. At the same time, users also pointed out some shortcomings of Kimi: coding capabilities and creative writing are still far behind ChatGPT and Claude, especially in complex programming tasks and creative content generation; the free version has usage restrictions and may require waiting in line during peak periods; although multi-modal capabilities are improving, they still need to be strengthened.
Judging from industry media and evaluation reports, Kimi has a high status in the domestic large model market. A Zhihu column review article stated that Kimi K2.5 "challenges Gemini 3" on multiple indicators, showing Dark of the Moon's ambition to benchmark against top international models. An analysis article from Tencent Cloud Developer Community pointed out that Kimi K2.6 outperformed GPT-5.4 and Claude Opus 4.6 in programming benchmark tests, marking the progress of domestic programming models. The financing progress of Dark Side of the Moon has also attracted industry attention. The approximately US$2 billion in financing completed in May 2026 will bring the company's valuation to more than US$20 billion, making it the most highly valued AI unicorn in the country. Meituan Longzhu led the investment, with participation from Hillhouse Capital, Tencent Investment and others, which shows the Internet giant’s layout for the AI track.
While Kimi and the Dark Side of the Moon is going well, it's also faced some controversy and potential risks. First of all, as an AI-generated content tool, the content generated by Kimi may contain factual errors, and users need to identify them themselves. The official website also reminds that "the content is generated by AI, please review it carefully." Secondly, as the number of users increases, the service quality of the free version may be affected, and response delays may occur during peak periods. In addition, the domestic large model market is highly competitive, and Kimi needs to continue investing in research and development to maintain a competitive advantage.
Kimi is suitable for the following groups: researchers, students and professionals who need to process long documents, such as reading papers, financial reports, reports and other long texts; domestic users who need a Chinese AI assistant, Kimi does not need to circumvent the firewall and has strong Chinese understanding ability; searchers who need to quickly obtain information answers, Kimi responds quickly and outputs concisely; for light AI users, the free version can meet their basic needs. Kimi is not suitable for the following people: professional programmers, it is recommended to use Claude or GPT when they need more powerful coding capabilities; users who need image generation function, Kimi does not currently support image generation; users who need highly creative writing, the creative content generation ability needs to be improved. For users who want to try Kimi, it is recommended to start with the free version and experience the basic functions before deciding whether to upgrade to the professional version. The professional version is suitable for users who need to process long documents or use AI frequently.
Kimi is the leading AI intelligent assistant in China, with ultra-long context and Chinese understanding capabilities as its core advantages. After completing a large amount of financing in 2026, the valuation will exceed US$20 billion and the development momentum will be strong. Kimi is suitable for domestic users who need long document processing, Chinese dialogue and information retrieval. Its free availability is an important advantage. However, coding capabilities and creative writing are still far behind the top international levels and are not suitable for users who need these functions. Overall, Kimi is an important choice in the domestic AI market and worth trying.

User Reviews

MargieHale

—

K3的降价策略挺聪明，输出从100元降到30元，直接拉低了使用门槛。之前嫌贵不敢用的开发者这下可以上了，不过刚充值的同学估计心态要崩了。用这个价格对标Fable 5，月之暗面这是确立高端定位了。用这个价格对标Fable 5，月之暗面这是确立高端定位了，勇气可嘉。

Christopher.PerezK

—

K3发布到现在不到十天就降价了，月之暗面是真的想抢开发者市场。不过老用户会不会觉得被背刺啊，昨天刚充了会员的同学心态估计已经炸了。早买早享受，晚买享折扣，千古不变的道理。

Dylan.Carter_88

—

7月25号降价消息出来了？昨天刚充的会员，心态有点崩……。不过能理解，毕竟2.8万亿参数的成本摆在那里。

Christian.Turner007

—

降价是好消息，但服务器扩容也得跟上啊。现在高峰期还是卡得让人抓狂，有钱没地方花才是最难受的。毕竟2.8万亿参数的推理成本不是闹着玩的，理解归理解，体验还是要优化。

AnvaJensen

—

终于体验到K3了，前端代码能力确实猛，写了个Three.js的3D场景，一次成型没改过。但跑一次烧掉我2万token，真心疼，会员额度根本扛不住这种消耗速度。如果速度问题能优化到K2.6的水平，那就是真正的完全体了。

DrLeoHall_2024

—

K3是真的好，也是真的贵。输出100元/百万token，国内模型里最贵的了吧？。毕竟是在国内模型里第一个敢定这个价的，勇气可嘉。

BlockKi_ng

—

K3断售这事儿真的离谱。发布才两天就叫停了新订阅，说是算力不够用。做AI产品��怕的就是用户来了你用不了，希望月之暗面赶紧把扩容搞完吧。毕竟这次算力缺口是短期问题，长期看好K3的潜力。毕竟这次算力缺口是短期问题，长期看好K3的潜力，希望产能快点跟上。

BillyColemanSr

—

用K3写了个react组件，确实比K2.6强很多，能理解复杂的业务逻辑。但响应速度慢得想砸键盘，一个小问题卡半天，开发节奏全被打乱了。目前这个响应速度要是不在后台跑长任务，交互体验真的不太行。目前这个响应速度要是不在后台跑长任务，交互体验真的不太行，急需优化。

梅花_6

—

用K3跑了三小时的代码任务，中间出了两次错自己修好了。虽然慢是慢了点，但能一直跑不崩这点值得肯定。虽然慢，但这正是Agent该有的样子——自己能修bug自己迭代。

Madison_White168

—

杨植麟回应K3争议那��采访看了，说算力缺口要几个月才能补上。感觉月之暗面这波是被自己的热度反噬了，希望别像其他国产AI一样高开低走。之前爆火的时候服务跟不上，热度过了再扩容就晚了。之前爆火的时候服务跟不上，热度过了再扩容就晚了，时间窗口很关键。

MatthewHughes_662

—

K3的前端能力确实没得黑，前端代码榜登顶不是吹的。但我后端任务跑了一下，跟DeepSeek V4 Pro差不多，没有拉开差距，偏科问题还是比较明显的。如果后端也能有前端这个水平就完美了。如果后端也能有前端这个水平就完美了，期待K3后续版本补齐短板。

天涯

—

用K3写了个网页游戏demo，我承认有被惊艳到。提示词就一句话，15分钟出完整可运行工程。但会员额度直接烧了三分之一……。量力而行吧，毕竟不是谁都有预算无限烧的。

xFélixGonzález_2024

—

为什么没人吐槽K3的幻觉率？从39%涨到51%，答对的题多了但编的也多了。用在合同审查这种场景谁敢放心啊，万一漏掉重要条款就麻烦了。只要这个幻觉问题能控制到20%以内，我马上就全面迁移过来。只要这个幻觉问题能控制到20%以内，我马上就全面迁移过来用K3。

goldenpanda348

—

我拿K3和Fable 5做了同样的代码重��任务，K3花了12分钟，Fable 5只要3.5分钟。时间差太悬殊了，如果赶工期的话谁等得起这么久啊。如果能把速度提到正常水平，这个价格完全可以接受。如果能把速度提到正常水平，12分钟缩到5分钟以内，这个价格就完全可以接受了。

Daniel.Bell_Plus

—

K3走开源路线真的聪明。2.8万亿参数直接开源，以后企业可以本地部署不用被闭源绑架。光这一点就值了。尤其对数据敏感的企业来说，本地化部署的价值远超价格本身。

Bryan30

—

作为一个前端开发者，K3确实让我有点想从Claude转过来。审美这块K3拿捏得死死的，生成的UI比GPT-5.6 Sol好看太多了，而且开源的特性让人很放心。Claude那种细腻的设计感K3学了个七七八八，非常难得。

CDiazSr

—

马斯克都说K3 impressive了，国产AI这波真长脸。虽然自己还没用上（断售了），但光看评测就已经很提气了。国内AI终于有了拿得出手的和国际巨头比拼的东西了。

CoryMarshall

—

K3的分析报告生成能力真是一绝。丢了三份券商研报进去，自动做了交叉对比，连数据口径的差异都标出来了。分析师表示感动，这活儿以前怎么也得干半天。特别是那种需要深度理解业务数据再做交叉验证的场景，K3真的无敌。

Emily.Jimenez_66

—

K3答问题真的好啰嗦啊……问它今天天气怎么样，给我写了八百字的分析报告还带数据可视化。我要的是天气不是天气预报论文好伐，有时候简洁点比什么都强。现在每次用K3之前都要想想是不是真的需要深度推理，不然用K2.6就够用了。

AlphaWolf35

—

用K3跑了一段视频剪辑任务，前后快两个小时。结果确实可以，但时间成本太离谱了，普通用户根本耗不起这个等待时间。做那种需要长时间连续推理的复杂任务，K3目前还是不可替代的。

zrky36

—

K3的定价真的迷，输出100元/百万token，比GLM-5.2贵了三倍多。你说能力强我认，但贵这么多能换来多高的成功率？如果任务没跑通，这钱不就白花了嘛。

AmberEvansK

—

今天测了K3的Kimi Claw，部署了一个定时抓取脚本。不用写代码一键部署，24小时在线自动跑，这个体验确实好。配合5000+社区技能库，能做的东西太多了。

Brenda.Martinez_77

—

我是Kimi��老用户了，从K2一路用到K3。这次K3虽然强，但免费版明显被削弱了，长文档上传动不动就乱码，关键条款经常识别不全。如果继续这么搞下去，忠实用户迟早会被逼走。这个吃相真的有点难看了。

Joe_Mitchell168612

—

一句话：K3适合有耐心的开发者，不适合赶deadlines的打工人。慢是真的慢，强也是真的强，就看你的场景了。

AAort

—

K3的编程能力强但偏科太严重了。让它写个前端demo能让我高潮，让它写个稍微复杂点的后端逻辑就开始摆烂。

JOhal

—

K3一发布就全球第三，国产模型历史上头一回吧？虽然可能很快就被赶超了，但这一刻值得记录。

Helen_Moore_X

—

试了K3的视觉理解能力，把一张UI截图丢��去让它还原成代码，效果炸裂。布局、间距、颜色全都对上了，基本不用改就能直接用，前端工作效率直接起飞。

ViiviNiskanen

—

看了5亿token实测K3那篇文章，Vision in the Loop确实让Agent进化了一个台阶。做完代码自动截图检查再修改，这才是真正的迭代开发。以前生成前端代码还得手动刷新页面看效果，现在Agent自己就能看图改图了，效率提升是实打实的。

s8hgi

—

K3第一天发布我就冲会员了，结果根本挤不进去。服务器比春运火车票还难抢，有意思吗？

realPaigeByrd_88

—

K3发布了！2.8万亿参数开源！前端代码榜登顶！三连击直接把AI圈炸翻了。月之暗面牛逼。

Kimi

In-depth Report

User Reviews

Similar Products