(By You Yunting) Recently, the AI coding assistant Cursor, developed by the American startup Anysphere, released its new AI coding feature, Composer 2. However, users discovered that its underlying model was in fact Kimi K2.5, developed by Moonshot AI, leading to a controversy over Cursor’s failure to credit the model as required by the license. After public confirmation of Moonshot AI’s head of pre-training, even Elon Musk weighed in. While open-source LLMs (Large Language Models) are often free to use, their user agreements are not mere formalities. Today, let’s discuss the special clauses hidden in these user agreements and the consequences of non-compliance therewith.
I. Case Review
Cursor, the American AI coding assistant, recently released its “Composer 2” feature claiming to be powered by its self-developed model. However, users quickly discovered that its underlying model was actually Moonshot AI’s Kimi K2.5, which was subsequently confirmed by Moonshot AI’s head of pre-training on Twitter. Elon Musk later replied, “Yeah, it’s Kimi 2.5,” sparking heated debate in the open-source community. Moonshot AI’s user agreement explicitly requires commercial products with over $20 million in monthly revenue or 100 million monthly active users (MAU) to prominently display the “Kimi K2.5” brand on their user interfaces. Cursor’s monthly revenue is estimated at $167 million—roughly eight times that threshold.
Notably, on the same day the controversy escalated, the parties appeared to have reached a commercial understanding. Moonshot AI publicly posted: “We are proud to see Kimi-k2.5 provide the foundation.” Cursor’s co-founder later explained that among multiple base models they evaluated, Kimi K2.5 was the strongest, so they then conducted additional pre-training and four-scale reinforcement learning on top of it. At the same time, he acknowledged that failing to mention Kimi K2.5 in its launch blog post was a “miss”.
II. Special Clauses in LLM User Agreements
Kimi’s user agreement sets two thresholds: monthly revenue over $20 million or MAU over 100 million. These terms are relatively lenient, requiring only attribution rather than prior approval. In contrast, many similar agreements set only one threshold and many require prior approval.
1. User Scale Thresholds
Meta’s LLaMA 2 License Agreement dictates that if the MAU of the products or services made available by or for the licensee, is greater than 700 million MAU in the preceding calendar month, the licensee must request a license from Meta, which Meta may grant to the licensee in its sole discretion, and the licensee is not authorized to exercise any of the rights unless or until Meta otherwise expressly grants the licensee such rights. Furthermore, using the model to train competing models is also prohibited.
2. Revenue Thresholds
Stability AI’s Stable Diffusion, an AI image generation software, set no revenue limits in its early versions, but has adopted a dual-track licensing system in its versions 3.0 and 3.5: free for small users and paid for large enterprises. Specifically, for organizations with annual revenue exceeding $1 million, the free community license is void, and a Professional or Enterprise license must be purchased.
3. Specific Branding Requirements
Kimi K2.5 uses a modified MIT license. Its core term stipulates that if the model is used for any of the commercial products or services that have more than 100 million MAU, or more than 20 million US dollars in monthly revenue, the licensee shall prominently display “Kimi K2.5” on the user interface of such product or service.
4. Unilateral Termination Rights
Gemma, Google’s lightweight open-source LLM, features two unique mechanisms. The Gemma license explicitly allows commercial use without negotiation with Google, but Google reserves the right to terminate the license in case of breach and also forbids using Gemma to train competing models.
III. Why Include These Clauses if Breach Does Not Lead to Damages?
One may ask: the “cost of breach” seems low—for example, Cursor would not necessarily have to pay damages to Moonshot AI for failing to credit Kimi—why include such clauses? In fact, as illustrated above, these open-source AI licenses share common features: the retained trigger rights exert compliance pressure on large companies while reserving commercialization room for open-source LLM providers.
1. Scale-Based Trigger Clauses: The core logic of these clauses is to offer free access for small and medium-sized enterprises and researchers while establishing a threshold for large-scale commercial use, by which open-source LLM providers manage to identify clients capable of generating significant commercial value and prompt them to pay for commercial licenses once they reach a certain scale. While LLM providers do not need to pursue every minor breach, “selective enforcement” against leading companies in the industry will bind them to meet compliance requirements. The recent Cursor incident is a prime example: exposure caused immediate deterrence, leading its co-founder to promptly acknowledge the issue which would have created material compliance risks for future financing and IPO due diligence.
2. Brand and Ecosystem Control: Kimi’s clauses focus on brand exposure. It does not demand payment, but mandates prominent attribution within the product interface to drive traffic. This strategy aims to rapidly elevate brand recognition and cultivate an ecosystem through widespread derivative apps. Free open-sourcing is often a phased strategy for LLM providers; once the brand is established and maintains technological leadership, its subsequent commercialization pathways become highly diversified with revenue generated from API and cloud service fees, private deployments customized for enterprises, and even ad-supported models currently being explored by OpenAI.
Ultimately, returning to the Cursor incident, its swift acknowledgement of the oversight and subsequent achievement of cooperation precisely demonstrates the practical effectiveness of such license clauses. For AI companies, while benefiting from open-source models, it is essential to re-examine the special clauses in user agreements, as the “open-source feast” is not truly free.
Lawyer Contacts
Disclaimer of Bridge IP Law Commentary
Short Link: