Coding Editor Cursor Weighs Dropping Downloadable Chinese AI Models

The maker of the Cursor coding editor plans to explore dropping Chinese AI models it can download and adapt. Paying users can still choose Composer 2.5, a coding model built on one of them.

TL;DR
  • Model Phaseout: Cursor’s maker is weighing whether to drop Chinese open-weight AI models it can download and adapt for its coding tools.
  • Composer Access: Paid Cursor users can still choose Composer 2.5, built on Kimi K2.5, a downloadable model from China’s Moonshot AI.
  • Federal Contracts: Two House committees want Congress to restrict Chinese-developed AI models in federal systems and in work on federal contracts.

Anysphere, the company behind the AI coding editor Cursor, which was aqcuired by SpaceX, told two House committees it plans to explore phasing out Chinese-developed open-weight models because of geopolitical risk. Developers on paid Cursor plans can still choose a coding model built on one of those downloadable models, with no phaseout timetable in the committees’ account.

The House Homeland Security Committee and the House Select Committee on China disclosed Anysphere’s position in a November 20 report, citing a briefing with company staff. In April the committees asked Anysphere how it chose its models and what contact it had with the companies that made them.

How Cursor Built Composer on Kimi

Moonshot AI released the weights for Kimi K2.5, the values a model learns during training. With those weights, another company can run the model on infrastructure it chooses and train it further for its own tasks. Cursor used that starting point for Composer 2, a model trained to write and edit code inside the editor. It kept the same base for Composer 2.5, which Cursor says follows complex instructions more reliably and stays with longer coding jobs than Composer 2.

Composer 2.5 and Grok 4.7 were both listed for paying Cursor users on November 21. Grok 4.7, developed by Cursor and SpaceXAI, which aqcuired Anysalso handles long coding tasks and builds on the earlier Grok 4.6. Cursor said in May that it was also training a larger model from scratch with SpaceXAI. The committees’ report names no replacement for Composer 2.5.

What Companies Control When They Run a Model

Running downloaded weights lets a company process requests without sending them to the original developer’s servers. The committees found that the companies they examined largely avoided passing data to Chinese-operated services by running models on their own infrastructure or through U.S. providers. The weights still carry the results of the original training, and a company building on them can depend on the developer for new versions or fixes. Cursor says it uses no infrastructure in China and uses no direct data processors headquartered there.

DoorDash used a Moonshot model for an initial review of software changes. It told the committees it ran its Chinese-developed models on U.S. infrastructure and blocked connections to China-based AI services from company devices and hosting systems. DoorDash kept the Moonshot model’s deployment, access permissions and monitoring in place after the White House accused Moonshot of improperly copying capabilities from a U.S. model and obtaining restricted chips. DoorDash told investigators the public disclosures gave it too little information to identify a security risk specific to its deployment.

DoorDash gave the committees a separate estimate for replacing one Chinese-developed model embedded in a larger system. That could take up to six months, while DoorDash could switch models in most other uses almost immediately.

What the House Committees Want to Change

An existing Department of Homeland Security order bars federal civilian agencies from buying information technology products and services developed by DeepSeek or its affiliates. It also orders those agencies to remove those offerings, along with third-party products and services built with DeepSeek, from their IT systems. The two House committees want Congress to broaden restrictions to Chinese-developed models and their derivatives in federal systems and to cover federal contract work.

The committee chairs also called for federal computing resources and agency use to help U.S. developers build open-weight alternatives. They argued that government agencies using those models could make private customers more willing to adopt them.

Markus Kasanmascheff
Markus Kasanmascheff
Markus has been covering the tech industry for more than 15 years. He is holding a Master´s degree in International Economics and is the founder and managing editor of Winbuzzer.com.
Subscribe
Notify of
guest
0 Comments
Newest
Oldest Most Voted