Cursor Faces Backlash Over AI Model Origin

Cursor Faces Backlash Over AI Model Origin

Cursor, a rapidly growing AI coding tool, recently launched its new AI model, Composer 2. The company announced that Composer 2 offers “frontier level intelligence at a low cost.” Early users were impressed, calling the model exceptionally good for coding tasks and significantly cheaper than other top-tier models.

However, the excitement quickly turned into controversy when a developer named Finn pointed out a potential issue. He suggested that Composer 2 was, in fact, a fine-tuned version of an open-source model called Kimmy K2.5, developed by the Chinese company Moonshot AI (also known as Kimi AI). Finn claimed that Composer 2 was essentially Kimmy K2.5 with added reinforcement learning, and that Cursor had not properly disclosed this.

What is Cursor and Kimmy K2.5?

Cursor is an AI-powered code editor built on top of VS Code, a popular open-source project. The company has seen massive growth, reportedly raising significant funding and achieving a high valuation. It is considered one of the most popular and fastest-growing AI coding applications available.

Kimmy K2.5 is a powerful open-source AI model from China-based Moonshot AI. It is known for its capabilities, especially in agentic tasks. While open-source, its license has specific requirements for larger companies. Businesses with over 100 million monthly active users or $20 million in monthly revenue must publicly disclose their use of the Kimmy K2.5 model within their products.

The Allegations and Cursor’s Response

The core of the controversy lies in Cursor’s initial lack of explicit mention of Kimmy K2.5 as the base model for Composer 2. Finn’s post, which gained traction on social media, highlighted evidence suggesting the connection. This led to further scrutiny, with even Elon Musk weighing in on the situation.

In response, Lee Robinson, who works for Cursor, stated that Composer 2 started from an open-source base and that Cursor planned to do full pre-training in the future. He mentioned that only about a quarter of the compute used for the final model came from the base, with the rest being Cursor’s own training, including reinforcement learning using their extensive user data. However, he avoided naming Kimmy K2.5 directly, referring only to an “open-source base.” He also indicated that Cursor was following the license terms through their inference partner.

Further adding to the confusion, a post from Yulun Du, who works at Moonshot AI, initially appeared to confirm the use of Kimmy K2.5. He reportedly stated that the tokenizer was the same as Kimmy’s and expressed shock that Cursor had not respected the license or paid fees. This post was later deleted.

Moonshot AI then released an official statement congratulating Cursor on the launch of Composer 2 and expressing pride that Kimmy K2.5 provided the foundation. They noted that seeing their model integrated through Cursor’s pre-training and reinforcement learning was part of the open model ecosystem they aimed to support. The statement also mentioned that Cursor accessed Kimmy K2.5 via Fireworks AI, an inference platform, as part of an authorized commercial partnership.

Why the Secrecy?

Cursor’s decision not to immediately disclose the use of Kimmy K2.5 appears to stem from several factors. Firstly, as a company valued at nearly $30 billion, presenting a model as entirely in-house rather than built upon another’s work can be seen as more impressive and aligns with their image as an AI research leader.

Secondly, the geopolitical climate surrounding US-China technology relations likely played a significant role. Disclosing that their flagship AI coding tool was based on a Chinese model could have created a public relations headache, especially for an American company seeking enterprise clients who are often sensitive to such ties.

Why This Matters

This incident highlights the complex realities of the AI development landscape. While open-source models like Kimmy K2.5 are intended to foster innovation, companies often build upon them with significant added value through their own data and training. The challenge lies in transparently attributing these contributions while also showcasing a company’s unique technological advancements.

Cursor argues that three-quarters of the compute and their extensive reinforcement learning efforts represent a substantial contribution beyond the original Kimmy K2.5 base. They also published a detailed blog post explaining their innovative approach to training, including a technique called “self-summarization.” This method allows the AI to pause and condense its progress, enabling it to handle much longer and more complex tasks than its context window would normally allow. This technique was demonstrated by its success in solving a challenging benchmark involving making the original Doom game run using an old computer language called MIPS.

The situation also raises questions about the broader open-source AI ecosystem. Clement Delang, founder of Hugging Face, commented that open source continues to be a major enabler of competition, validating the significant role of Chinese open-source contributions. He suggested that the frontier of AI is now less about who trains models from scratch and more about who can adapt, fine-tune, and commercialize them most effectively.

Ultimately, while Cursor faced criticism for not being upfront about the origins of Composer 2, the company appears to have followed licensing terms through its partnership with Fireworks AI. The incident underscores the ongoing debate about attribution in AI development and the strategic considerations companies face when integrating open-source technologies, particularly in a globally competitive and politically sensitive tech environment.

Source: Cursor is CAUGHT red handed… (YouTube)