Do Shell-Based Chinese Large Models Support a $5 Billion Valuation? Uncovering Cursor's "Shell Company" Controversy

1、

This past weekend saw the most sensational AI news that shook both sides of the Pacific, and even caught Elon Musk's attention.

Yesterday, AI programming tool Cursor released its "own" model Composer 2.

The image above is a screenshot from their official website, which still states "proprietary model" when you click on it.

Since Composer 1 was released in October 2024, the outside world has been suspicious that it's a wrapper around a Chinese model because its behavior is very similar, but they lacked concrete evidence.

Now that Composer 2 has arrived, many people have started investigating what model actually powers it. Is it really Cursor's own proprietary model?

Cursor implemented many restrictions to prevent reverse-engineering, but there was a small oversight. A foreign Twitter user @fynnso discovered that there's one place that was restricted in the previous version but is executable in this version.

First, you set up your own server to act as the interface for calling AI models. Whether you actually have a model doesn't matter—you just need to be able to receive client requests.

Then, in your local Cursor, you configure the model to use as Composer 2, with the model URL pointing to the server you just set up. This way, Cursor will send requests to your server, allowing you to see exactly what model it's requesting.

The truth was exposed: the model ID it was requesting turned out to be kimi-k2p5-rl-0317-s515-fast (see image below).

2、

This foreign Twitter user then posted the screenshot online. That's when things exploded. Astute observers immediately recognized this as solid evidence: Composer 2 is actually a wrapper around Kimi K2.5.

What's ironic is that as soon as the scandal broke, Cursor immediately patched the loophole, and now this request can no longer be reproduced (see image below).

But it was too late. The information spread across the internet, and even Elon Musk tweeted: "It's just Kimi K2.5".

Now it had become an open secret that could no longer be hidden.

3、

People's attention quickly shifted to whether Cursor had violated any licensing terms. Although Kimi K2.5 is an open-source model, it uses a modified MIT license (see image below).

The license states: you can use this model freely, with one condition—if your commercial product has over 100 million monthly active users, or monthly revenue exceeding $20 million, you must clearly disclose in your user interface that you are using Kimi K2.5.

Cursor's most recent disclosed annual revenue is $2 billion, equivalent to approximately $167 million in monthly revenue, which clearly meets the conditions above. However, it concealed the fact that it was using K2.5.

Just when everyone had determined that Cursor was infringing on the license, one of their leaders finally couldn't sit still and spoke up.

He admitted they indeed use Kimi K2.5, but denied any infringement, stating that their license comes from their partner Fireworks AI.

Shortly after, Kimi's official account also tweeted.

Kimi's official account confirmed that Cursor obtained authorization from Fireworks AI, a Silicon Valley AI company founded by Chinese entrepreneurs. Fireworks AI specializes in fine-tuning AI models and reinforcement learning. They obtained authorization from Kimi to further train the model, and then granted a sublicense to Cursor.

4、

At this point, the situation is mostly clear. Cursor did not violate Kimi's licensing terms, so there is no infringement.

If that's the case, why did it go to such great lengths to cover up this fact? Would it have been so difficult to openly acknowledge that it provides a modified version of Kimi K2.5?

I suspect the reason is related to Cursor's rapidly inflating valuation.

Bloomberg reported this month that Cursor is in the process of raising its next round of funding, with a valuation of $50 billion.

Do you know what its valuation was before?

In October 2023, when Cursor was founded, its valuation was $50 million. In August 2024, its Series A valuation rose to $400 million. In December, its Series B valuation jumped to $2.6 billion. In its latest round in November 2025, the valuation had already reached $29.3 billion.

As you can see, every few months, the valuation doubles. This rocket-speed appreciation rate needs performance to back it up. But Cursor itself is just a modified version of VS Code, using only open-source technology.

To support increasingly high valuations, it has motivation to rebrand itself from an AI tool into a large language model company with model R&D capabilities.

I believe this is the main reason it was unwilling to disclose that it uses Kimi K2.5.

5、

Looking at the entire incident, Cursor is undoubtedly the loser, while Kimi is the winner, gaining enormous high-value exposure for free.

When Cursor released Composer 2, it disclosed performance and cost comparisons.

Composer 2's performance 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 its cost is also the lowest.

Since Composer 2 is essentially a fine-tuned version of Kimi K2.5, then using Kimi directly would achieve the same results.

6、

Historically, people abroad have always accused Chinese companies of stealing foreign technology. But this incident proves that Chinese companies also have technology to export. Those celebrated foreign companies are also quietly using Chinese technology behind the scenes.

Recalling last week, Kimi's founder Yang Zhilin received an invitation from Jensen Huang to speak at the Nvidia GTC conference, as the sole representative of a Chinese large language model company.

On stage, he presented a paper that the Kimi team had just published: Attention Residuals.

This new technology is said to significantly enhance the inference capabilities of large language models.

My thought is that we should have confidence in domestic large language models. You can rest assured using them in your daily work. The gap between domestic large language models and flagship international models is constantly shrinking, and the prices are reasonable.

According to Yang Zhilin, the next K3 model to be released will have massive performance improvements, at least 10 times better than K2.5. We can look forward to that.

（End）

Document Information

• Copyright Notice: Free to share—non-commercial—non-derivative—attribution required (Creative Commons 3.0 License)
• Publication Date: March 21, 2026

Further Reading:

• The River of Art: Exploring Cultural Treasures Sealed by Time
• Technology Enthusiast Weekly (Issue 390): Without Training Data, Large Language Models Are Idiots

This article was automatically generated and optimized by an intelligent content system for reference only.