1.
This weekend, the most buzzed-about AI news shook both sides of the Pacific, even catching Elon Musk's attention.
Yesterday, the AI programming tool Cursor launched its "own" model Composer 2.
The screenshot above is from the official website, and it still says "Proprietary Model" when you click through.
Since Composer 1 was released in October 2024, there has been suspicion that it was a Chinese model in disguise, due to its similar behavior, but evidence has been hard to come by.
Now that Composer 2 has arrived, many are investigating what model lies behind it—is it truly Cursor's own creation?
Cursor implemented many restrictions to prevent hacking, but no plan is perfect. The foreign Twitter user @fynnso discoveredThere is a place that was forbidden in the previous version but can be executed in this version.
First, set up a server yourself to act as the interface for calling the AI model. It doesn't matter which model you use, as long as it can receive client requests.
Then, in your local Cursor, set the model to Composer 2, and the model URL is the server you just set up. This way, Cursor will send a request to your server, allowing you to see exactly what model it is requesting.
The truth came to light; the model ID it requested was actually kimi-k2p5-rl-0317-s515-fast (as shown in the image below).
2,
This foreign Twitter user then took the screenshot above,PublishOnline. Now this blew up, the discerning people could see, this is solid evidence.Composer 2 is actually a wrapper for Kimi K2.5.
Ironically, as soon as the issue blew up, Cursor quickly patched the vulnerability, and now this request can no longer be reproduced (see image below).
But it was too late, as the news had already spread online, and even Elon Musktweeted:"It's just Kimi K2.5."
Now, it's become an open secret and can no longer be hidden.
3,
Everyone's attention quickly shifted to whether Cursor was infringing, because while Kimi K2.5 is an open-source model, it uses a modified MIT license (see image below).
License states: You are permitted to use this model freely, with the sole condition that if your commercial product has over 100 million monthly active users or generates over $20 million in monthly revenue, you must disclose prominently in the user interface that you are using Kimi K2.5 .
Cursor's recently disclosed annual revenue is $2 billion, equivalent to $167 million per month, clearly meeting the above conditions. However, it concealed the fact that it uses K2.5.
Just as people were concluding that Cursor was infringing, one of their executives finally couldn't sit still and came forward to speak
. He admitted that they indeed use Kimi K2.5 but claimed no infringement, stating their license comes from partner Fireworks AI.
Later, the official Kimi team alsotweeted.
Kimi officially confirmed that Cursor obtained authorization from Fireworks AI. The latter is a Chinese AI company in Silicon Valley that specializes in fine-tuning AI models and reinforcement learning. It obtained authorization from Kimi to retrain the model, and then licensed it to Cursor.
4,
the matter is basically clear at this point. Cursor did not violate Kimi's authorization terms, so there is no infringement.
Given that, why is it desperately trying to hide the fact? Is it difficult to openly admit and provide the modified Kimi K2.5 model?
I speculate that, the reason is related to Cursor's continuously growing valuation.
Bloomberg reported this month that Cursor is conducting its next round of funding, with a valuation reaching $50 billion.
You know what? What was its previous valuation?
In October 2023, when Cursor was founded, its valuation was $50 million; in August 2024, during its A round of funding, the valuation rose to $400 million; in December, during its B round of funding, the valuation rapidly increased to $2.6 billion; and in November 2025, the latest round of funding saw the valuation reach $29.3 billion.
It can be seen that every few months, the valuation doubles. This rocket-like growth rate requires performance support. But it is essentially just a modified version of VS Code, using all open-source technologies.
To support the increasingly high valuation, it has the motivation to position itself as a large model company capable of model development, rather than just an AI tool.
I think this is the main reason why it is unwilling to disclose the use of Kimi K2.5.
5,
Throughout the entire event, Cursor is undoubtedly the loser, while Kimi is the winner of this time, getting a large amount of high-value exposure for free.
Cursor released Composer 2, disclosing performance and cost comparisons.
The performance of Composer 2 is lower than GPT-5.4 but higher than Opus 4.6.
However, its generation speed is faster than both GPT-5.4 and Opus 4.6, and the cost is also the lowest.
Since Composer 2 is just a fine-tuned version of Kimi K2.5, thenYou can also get the same effect by directly using Kimi..
6,
In the past, there were always people abroad accusing Chinese companies of stealing foreign technology. However, this incident has proven that Chinese companies also have technology exports. Those famous foreign companies, in secret, are also sneakily using Chinese technology.
Recalling last week, Yang Zilin, founder of Kimi, received an invitation from Huang Renxun to attend the Nvidia GTC conferenceSpeechIt is the only representative of a Chinese large model company.
He was speaking on stage about the paper just published by the Kimi teamAttention Residual(Attention Residuals)
This new technology is said to significantly improve the reasoning capabilities of large models.
My idea is that everyone should have confidence in domestic large models, and they can be used with complete peace of mind in daily work. The gap between domestic large models and overseas flagship models is constantly narrowing, and they are also affordable.
According to Yang Zhilin, the next K3 model to be released will have significant performance improvements, even if it's not 10 times stronger, it will be much better than the K2.5, and we can look forward to it.
(End)