⌘K
Change language Switch ThemeSign In
Narrow Mode
惯犯 Cursor 又来"自研"了，还是熟悉的配方，国产模型套壳
刘 刘小排r @刘小排
One Sentence Summary
This article reveals that AI coding tool Cursor's claimed self-developed Composer 2 model is actually fine-tuned from Moonshot AI's open-source Kimi K2.5 model, and fails to credit the source as required by the open-source license, criticizing its hypocritical behavior.
Summary
The article provides a detailed report on yet another "self-development" farce by Silicon Valley AI coding tool Cursor. On March 19, 2026, Cursor highly publicized the release of Composer 2, claiming it was developed through continuous pre-training and large-scale reinforcement learning, with cost-performance that dwarfed Claude and GPT. However, just 3 hours later, developers discovered its actual model ID was kimi-k2p5-rl-0317-s515-fast, meaning Composer 2 was actually fine-tuned from Moonshot AI's open-source Kimi K2.5 model. The Kimi team confirmed the tokenizer was completely identical and questioned why Cursor didn't comply with the open-source license (which requires prominent attribution of Kimi when monthly revenue exceeds $20 million). The article also mentions that Cursor's previous generation Composer 1 was also found to be based on Alibaba's Tongyi Qianwen, and Windsurf was discovered to be using Zhihu AI's GLM model at its core. The author criticized this behavior of using open-source models while pretending to have self-developed them and violating open-source licenses as "hypocritical," lamenting that the wheel has turned.
Main Points
* 1. Cursor Composer 2 was confirmed to be fine-tuned from the open-source Kimi K2.5 model, with the actual model ID exposing the underlying model source.Developers discovered that Composer 2's API request contained the model ID kimi-k2p5-rl-0317-s515-fast, and the Kimi team confirmed the tokenizer was completely identical, essentially confirming the wrapper behavior. * 2. Cursor is suspected of violating the Kimi K2.5 open-source license by failing to attribute Kimi as required when monthly revenue exceeds $20 million.Cursor's parent company Anysphere has a valuation of $29 billion, annual revenue exceeding $2 billion, and monthly revenue of approximately $167 million—over 8 times the license threshold—but the interface shows no Kimi attribution whatsoever. * 3. This isn't the first time Cursor was caught using an open-source model—Composer 1 was also found to be based on Alibaba's Tongyi Qianwen.Users discovered occasional Chinese characters in code generated by Cursor, sparking community doubts. Cursor's researcher only responded that the focus was on "RL post-training," without directly denying it. * 4. There's nothing wrong with using open-source models themselves, but pretending to have self-developed them, violating open-source licenses, and deliberately hiding sources are acts of industry dishonesty.Using open-source models for the world is a good thing, but one should openly acknowledge it and attribute sources per the license requirements, rather than packaging it as self-developed and removing all traces. * 5. Chinese AI open-source models are gaining global recognition, with the shift from "Chinese self-developed" to "American self-developed" reflecting China's growing technological strength.From ridiculing Chinese companies for repackaging open-source models to Silicon Valley companies now building products based on Chinese open-source models, the shifting tide reflects the progress of Chinese AI technology.
Metadata
AI Score
78
Website mp.weixin.qq.com
Published At Today
Length 1677 words (about 7 min)
Sign in to use highlight and note-taking features for a better reading experience. Sign in now
原创 刘小排 2026-03-21 13:27 北京
以前美国一开源，中国就"自研"。现在中国一开源，美国就"自研"。
哈喽，大家好，我是刘小排。
以前上学的时候，有句话叫"美国一开源，中国就自研"——讽刺的是中国公司拿开源模型换个皮就说自主可控。
现在风水轮流转了。中国一开源，美国就"自研"。
今天这事我又能笑一天——硅谷最火的AI编程工具Cursor，刚刚又双叒叕被抓了个大包。
怎么回事呢？
3月19日，Cursor高调发布了自家"自研"模型Composer 2。发布会上的说法是：我们做了持续预训练，又做了大规模强化学习，性价比吊打Claude Opus 4.6和GPT-5.4。
好家伙，一夜之间，硅谷科技媒体集体高潮：Cursor要跟OpenAI和Anthropic正面硬刚了！
!Image 10: 图片 仅仅过了不到3个小时，就翻车了。
一个开发者在调试Cursor的API请求时，发现了一个不得了的东西——Composer 2的真实模型ID是：
kimi-k2p5-rl-0317-s515-fast
你不需要懂技术也能读懂这串字符。kimi-k2p5，就是月之暗面的Kimi K2.5。rl，就是强化学习。
翻译成人话就是：所谓"自研"的Composer 2，就是拿了中国公司月之暗面的开源模型Kimi K2.5，做了一轮强化学习微调，然后换了个名字端上来的。
Kimi预训练负责人杜育伦直接在推特上@了Cursor联合创始人Michael Truell，说：我们测试了Composer 2的tokenizer，和Kimi的tokenizer完全一致。几乎可以确定，这就是我们的模型被进一步微调的结果。
然后杜总问了一个很直白的问题：为什么不遵守我们的许可证？为什么没有付费？ （注：杜总现在已经删除了这条）
这个问题问得好啊。
Kimi K2.5的开源许可证里写得清清楚楚：如果你的商业产品月收入超过2000万美元，必须在用户界面上显著展示"Kimi K2.5"字样。
Cursor的母公司Anysphere，估值290亿美金，年收入超过20亿美元——月收入大概1.67亿美元，是门槛的8倍以上。
但Cursor的界面上，只写了"Composer 2"。
一个字都没提Kimi。
可爱的马老师也发来了嘲讽
作为对比，Kimi官方Twitter显得非常大度和体面
你以为这是第一次？
Cursor上一代模型Composer 1，也被扒出来大概率是基于阿里的通义千问Qwen。证据是用户发现Cursor生成的代码里，时不时冒出中文字符。当时有人在社区问Cursor的研究员：你们的模型到底是不是基于开源模型微调的？
Cursor研究员的回答非常经典——他没说是，也没说不是，只说了一句："我们的重点在RL后训练上。"
哈哈哈哈，好一个"重点在RL后训练"。
就好比你问一个厨师："这道菜是不是用的预制菜？"他回答你："我们的重点在摆盘上。"
而且不止Cursor一家。
Windsurf——就是那个被Google DeepMind数十亿美金收购的编程工具——它的SWE-1.5模型也被发现底层其实是智谱AI的GLM模型。
作为曾经全世界科技创新灯塔的硅谷，现在已经纷纷痴迷于预制菜摆盘技术了吗？
其实我个人对"用开源模型做产品"这件事本身一点意见都没有——开源的意义就在于此，全世界都可以用，这是好事。 但你不能一边用着别人的模型，一边假装这是自己的，还违反开源协议、花那么多心思把人家名字全部去掉、藏起来。
这叫什么？这叫又当又立。
中国AI团队在开源社区贡献了Kimi K2.5、Qwen、DeepSeek、GLM这些优秀的模型，硅谷公司拿去用了，用了就用了，大大方方承认多好？在界面上写一行"Powered by Kimi K2.5"，既合规又体面，还能显得你有眼光。
结果呢你偷偷用了，还要包装成自研，被抓包了还不回应。 Cursor的评论区全是一种声音：“你敢不敢说Kimi两个字？” 经过两天的半推半就，Cursor现在终于算是勉强回应了，语气中中一万个不情愿
290亿美金的估值，连一句"谢谢"都难说出口。
这事儿啊，不丢Kimi的人，丢的是Cursor的人。
以前美国一开源，中国就"自研"。现在中国一开源，美国就"自研"。
时代真的变了。
期待你的想法，评论区见！ 跳转微信打开
刘 刘小排r @刘小排
One Sentence Summary
This article reveals that AI coding tool Cursor's claimed self-developed Composer 2 model is actually fine-tuned from Moonshot AI's open-source Kimi K2.5 model, and fails to credit the source as required by the open-source license, criticizing its hypocritical behavior.
Summary
The article provides a detailed report on yet another "self-development" farce by Silicon Valley AI coding tool Cursor. On March 19, 2026, Cursor highly publicized the release of Composer 2, claiming it was developed through continuous pre-training and large-scale reinforcement learning, with cost-performance that dwarfed Claude and GPT. However, just 3 hours later, developers discovered its actual model ID was kimi-k2p5-rl-0317-s515-fast, meaning Composer 2 was actually fine-tuned from Moonshot AI's open-source Kimi K2.5 model. The Kimi team confirmed the tokenizer was completely identical and questioned why Cursor didn't comply with the open-source license (which requires prominent attribution of Kimi when monthly revenue exceeds $20 million). The article also mentions that Cursor's previous generation Composer 1 was also found to be based on Alibaba's Tongyi Qianwen, and Windsurf was discovered to be using Zhihu AI's GLM model at its core. The author criticized this behavior of using open-source models while pretending to have self-developed them and violating open-source licenses as "hypocritical," lamenting that the wheel has turned.
Main Points
* 1. Cursor Composer 2 was confirmed to be fine-tuned from the open-source Kimi K2.5 model, with the actual model ID exposing the underlying model source.
Developers discovered that Composer 2's API request contained the model ID kimi-k2p5-rl-0317-s515-fast, and the Kimi team confirmed the tokenizer was completely identical, essentially confirming the wrapper behavior.
* 2. Cursor is suspected of violating the Kimi K2.5 open-source license by failing to attribute Kimi as required when monthly revenue exceeds $20 million.
Cursor's parent company Anysphere has a valuation of $29 billion, annual revenue exceeding $2 billion, and monthly revenue of approximately $167 million—over 8 times the license threshold—but the interface shows no Kimi attribution whatsoever.
* 3. This isn't the first time Cursor was caught using an open-source model—Composer 1 was also found to be based on Alibaba's Tongyi Qianwen.
Users discovered occasional Chinese characters in code generated by Cursor, sparking community doubts. Cursor's researcher only responded that the focus was on "RL post-training," without directly denying it.
* 4. There's nothing wrong with using open-source models themselves, but pretending to have self-developed them, violating open-source licenses, and deliberately hiding sources are acts of industry dishonesty.
Using open-source models for the world is a good thing, but one should openly acknowledge it and attribute sources per the license requirements, rather than packaging it as self-developed and removing all traces.
* 5. Chinese AI open-source models are gaining global recognition, with the shift from "Chinese self-developed" to "American self-developed" reflecting China's growing technological strength.
From ridiculing Chinese companies for repackaging open-source models to Silicon Valley companies now building products based on Chinese open-source models, the shifting tide reflects the progress of Chinese AI technology.
Key Quotes
* The so-called "self-developed" Composer 2 is essentially taking the open-source model Kimi K2.5 from Chinese company Moonshot AI, running one round of reinforcement learning fine-tuning, and serving it under a new name. * What do you call this? This is being hypocritical. * You can't use someone else's model while pretending it's your own, violating the open-source license, spending so much effort removing all traces of their name and hiding it. * The comments section on Cursor is filled with one voice: "Do you dare to say the word Kimi?" * Previously, when America open-sourced something, China would "self-develop." Now when China open-sources something, America "self-develops." The times have really changed.
AI Score
78
Website mp.weixin.qq.com
Published At Today
Length 1677 words (about 7 min)
Tags
Cursor
Kimi K2.5
Open Source License
AI Coding Tools
Model Wrapper
Related Articles
* Inside Cursor 2.4: Subagents Parallelism, Native Image Generation, and Code Traceability * Claude Code Browser Automation: Which Solution to Choose? within the Claude Code environment, providing selection recommendations.") * Exploring the Boundaries of AI Programming: Spec Coding Project Practice Based on Claude Code | Dewu Tech * Vol.104 | OpenClaw: From Installation to On-the-Job—It's Not the Model Getting Stuck Coding, It's Me in Front of the Screen * OpenCode Hands-on Review: Almost Useless for Top-Tier Models, But Why I Still Recommend It * Stop Checking Leaderboards! Ordinary People Can Also Measure the Real Gap Between Major Programming Models * Kimi Releases and Open-Sources K2.5 Model, Bringing New Visual Understanding, Code, and Agent Cluster Capabilities * Andrej Karpathy's 2025 Retrospective: We Are Summoning Ghosts! * One-Click Solution to Enhance LLMs' Frontend Aesthetic Capability * Dynamic Context Discovery HomeArticlesPodcastsVideosTweets