Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarTNW→ original

Cursor May Drop the Chinese Models Behind Composer 2

Cursor, the AI code editor, is considering dropping Chinese open-weight models over "geopolitical risk," according to US House committees. Its flagship Composer 2 grew out of Kimi K2.5, a model from Moonshot AI, which Cursor fine-tuned on code. There is no decision and no timeline yet. If your product runs on someone else's open model, plan a replacement in advance.

AI-processed from TNW; edited by Hamidun News
Cursor May Drop the Chinese Models Behind Composer 2
Source: TNW. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTNW ↗
Editor-in-Chief
DoorDashswapping one model in a large system can take up to 6 months
Key metric #2Airbnb runs Qwen-based models handling about 80,000 conversations a day and processing 250,000 support requests
Key metric #3SpaceX bought Cursor in August at an implied equity valuation of $60 billion

Imagine you built a product on someone else's foundation, and then regulators started asking questions about that foundation. That is roughly the position Cursor, the popular AI code editor, now finds itself in. Its flagship model, Composer 2, was built on Kimi K2.5, an open model from the Chinese company Moonshot AI. "Open" here means the developer published the model weights: they can be downloaded, run on your own hardware, and fine-tuned for specific tasks. Cursor did exactly that, taking Kimi K2.5 as a starting point and fine-tuning it on programming code. Now, US House committees say Anysphere, the company behind Cursor, plans to study moving away from Chinese-developed models because of geopolitical risk.

What Cursor Actually Said

The report was prepared by the House Committee on Homeland Security and the Select Committee on the CCP. According to them, Anysphere itself told them about this plan. But the wording is cautious: the company "plans to study" a gradual phase-out rather than announcing one. The report contains no decision and no deadline. Cursor, SpaceX, and DoorDash did not respond to CNBC's request for comment, and Airbnb declined to comment.

The context matters. SpaceX completed its acquisition of Cursor in August, at an implied equity valuation of $60 billion, according to CNBC. The committees note that the Chinese foundation of Composer 2 became public only after a third-party developer spotted it. In Cursor's own description, Kimi K2.5 is now called a starting point.

How Other Companies Use Chinese Models

The investigation began in April with letters to Anysphere and Airbnb, and DoorDash was added in July. It turned out that all three companies use Chinese models in different ways.

Airbnb runs adapted versions of Alibaba's Qwen on American servers. One version powers an assistant that handles about 80,000 conversations a day with guests and hosts. Another summarizes roughly 250,000 support tickets and calls each day.

DoorDash counted 25 families of Chinese models in production, and each costs at least $5,000 to replace. If a model is embedded in a large system, swapping it can take up to six months. The company says most other use cases can switch almost immediately. The difference is easy to explain: where a model is wired deep into a system, moving it becomes a project, while where it sits behind a clean interface, it can simply be replaced with another.

Why Companies Choose Chinese Models, and What Lawmakers Propose

The report's authors cite three reasons: price, the ability to fine-tune, and quality close to the frontier. In their assessment, the US has no open model with a suitable combination of price and capability. They also write that a significant part of China's progress relies on capabilities drawn from American models: in September, OpenAI said it had detected attempts to extract reasoning skills from its AI. Hosting a model on American servers keeps user data outside China, but it does not change how that model was trained.

The committees propose to:

  • extend existing federal restrictions to other Chinese developers;
  • bar their models from federal contracts and from high-risk critical infrastructure;
  • create a model provenance standard;
  • launch a national American open-model program.

They also warn against haste. If companies are forced off Chinese models before reliable alternatives exist, costs will rise and the market will fall to a few closed providers. The authors point to new open models from Reflection AI and Mistral as encouraging signs.

What Developers and Product Managers Should Do

The report does not directly ban anything, but it sets a direction. If you work with government clients or critical infrastructure, find out which base model sits under your product, including fine-tuned derivatives. Keep an internal model registry that records origin, since a provenance standard will likely require exactly this kind of tracking.

The second takeaway is engineering. Put the model behind your own abstraction layer and maintain a test suite that lets you quickly compare a replacement. Then a switch takes days rather than six months, which is the worst case DoorDash described.

The limits are also clear. This is a policy document, not a law. Cursor has made no decision, and the data on model usage came from the companies themselves. But the signal is unmistakable: the Chinese origin of an open model is becoming a risk factor in technology selection, and it is worth building that into your plans now.

Story

Key Questions Answered

What Cursor Actually Said?

The report was prepared by the House Committee on Homeland Security and the Select Committee on the CCP. According to them, Anysphere itself told them about this plan. But the wording is cautious: the company "plans to study" a gradual phase-out rather than announcing one. The report contains no decision and no deadline. Cursor, SpaceX, and DoorDash did not respond to CNBC's request for comment, and Airbnb declined to comment. The context matters.

How Other Companies Use Chinese Models?

The investigation began in April with letters to Anysphere and Airbnb, and DoorDash was added in July. It turned out that all three companies use Chinese models in different ways. Airbnb runs adapted versions of Alibaba's Qwen on American servers. One version powers an assistant that handles about 80,000 conversations a day with guests and hosts. Another summarizes roughly 250,000 support tickets and calls each day.

Why Companies Choose Chinese Models, and What Lawmakers Propose?

The report's authors cite three reasons: price, the ability to fine-tune, and quality close to the frontier. In their assessment, the US has no open model with a suitable combination of price and capability. They also write that a significant part of China's progress relies on capabilities drawn from American models: in September, OpenAI said it had detected attempts to extract reasoning skills from its AI.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…