Cursor is looking for a way out of Chinese AI models, months after using one to build Composer 2. The SpaceX-owned coding platform told congressional investigators it plans to explore phasing out Chinese-developed open-weight models because of geopolitical risk.

The disclosure appears in a joint House report released November 20. It puts Cursor's model choices under a different kind of pressure: the Chinese technology that helped make its coding tool competitive is now part of a national security debate. The report gives no final decision or deadline for a switch.

Composer 2's Chinese foundation is documented

Cursor's developer, Anysphere, was one of three companies examined by the House Committee on Homeland Security and the House Select Committee on China. The investigation began in April and also covered Airbnb and DoorDash.

The report contrasts Anysphere's response with companies that had continued using Chinese models without changing their deployment arrangements:

By contrast, Anysphere informed the Committees that it plans to explore phasing out PRC-developed open-weight models because of geopolitical risk.

Cursor's own account explains how deeply a Chinese model was involved in Composer 2. The company launched the product on March 19, then published a technical explanation on March 27 identifying Moonshot AI's Kimi K2.5 as its base model.

That was the starting point for further training, rather than simply a Chinese service running under a different name. Cursor's explanation describes the work:

Composer 2 is trained in two phases: continued pretraining on a data mix that emphasizes code to deepen the base model's coding knowledge, followed by large-scale reinforcement learning to improve end-to-end agent performance.

The reinforcement learning took place in realistic Cursor sessions, using the same tools and setup available to the deployed model. The company also thanked the team behind Kimi K2.5 in its technical write-up.

Cheap, adaptable models are hard to replace

Open-weight models let businesses download trained parameters, run them on infrastructure they control and modify them for particular jobs. That flexibility is a major part of their appeal.

The congressional findings identify cost, customization and performance close to leading models as the main reasons American companies have adopted Chinese alternatives. They also conclude that the United States lacks a competitive open-weight option at the price and performance the market demands.

Replacing an embedded model can take work. DoorDash told investigators that replacing one Chinese-developed model inside a larger system could take up to six months, although most of its other uses could be switched almost immediately. That estimate concerns DoorDash, not a timetable for Cursor.

Cursor's original Composer 2 launch also made cost a selling point. It offered the standard version at $0.50 per million input tokens and $2.50 per million output tokens, alongside a faster, more expensive variant. Those were its March launch prices.

Congress wants restrictions and an American alternative

The chairmen's recommendations call for removing Chinese-developed models from federal systems, federal contract work and high-risk functions of critical infrastructure. They also propose standards documenting which base models and subsequent training steps sit inside an AI product.

Those are policy recommendations, rather than a new blanket ban created by the report. The lawmakers argue that running a model on American servers does not remove risks tied to its development or behavior under conditions ordinary testing might miss.

They also acknowledge a commercial problem with forcing companies to switch too quickly. Removing Chinese models before trusted alternatives exist could raise costs and concentrate business among a few closed-model providers. Their proposed American open-weight initiative would combine public computing resources, shared testing infrastructure and federal purchasing demand.

Cursor is already navigating a separate uncertainty over American models. OpenAI announced in August that it intended to wind down its model-supply contract following SpaceX's acquisition. A November 12 update says access was temporarily extended beyond the proposed shutoff date.

That leaves Cursor weighing a change to one source of model technology while another supplier relationship remains unsettled. Its next choice will have to account for the price, coding performance and freedom to customize that made an open Chinese base attractive in the first place.