Anysphere Considers Phasing Out Chinese Open-Weight Models in Cursor
Cursor developer Anysphere told U.S. House investigators that, due to geopolitical risks, it plans to explore phasing out open-weight models developed in China, but has not made a final decision or established a timeline. A joint congressional report said Cursor’s Composer 2 is based on Kimi K2.5 and underwent code-intensive continued pretraining and large-scale reinforcement learning. The report recommends establishing model provenance standards and expanding related restrictions, while warning that premature replacement could raise costs and deepen reliance on closed-source model providers.
Why read this
SelectedAI score 80
The report juxtaposes Cursor’s model sources and costs, replacement timeline, and regulatory recommendations, illustrating the implementation trade-offs involved in restricting Chinese open-weight models.
Topics
Body
Anysphere, the developer of the Cursor AI coding platform, told House investigators it plans to explore phasing out Chinese-developed open-weight AI models because of geopolitical risk, according to a joint congressional report released November 20, 2026.
The disclosure is an exploratory plan, not a completed phase-out, and the committees described no final decision or timetable. The House Committee on Homeland Security and the House Select Committee on China published the findings after an investigation that began in April 2026 and expanded from Airbnb and Anysphere to DoorDash on July 31, 2026.
Open-weight models have trained parameters that can be downloaded, run, and modified independently of the original developer, allowing derivatives to spread without a continuing commercial relationship. The committees said Composer 2, Cursor’s flagship proprietary model, was built on a Moonshot AI base whose identity became public only after an independent developer identified it.
Cursor’s technical report identifies that base as Kimi K2.5. Composer 2 underwent continued pretraining on a code-heavy data mix, followed by large-scale reinforcement learning in realistic Cursor sessions using the same tools and harness as the deployed model.
Why Chinese Models Became Embedded
Airbnb reported that customized versions of Alibaba’s Qwen models support an AI assistant handling roughly 80,000 guest and host interactions daily, alongside a fine-tuned model summarizing roughly 250,000 support cases and calls each day. Both run on U.S.-based service-provider infrastructure.
DoorDash reported 25 Chinese model families authorized and deployed in a non-trivial manner, defined as incurring at least $5,000 in computing and serving costs. It estimated that replacing one Chinese-developed model embedded in a larger system could take up to six months, although most other uses could transition almost immediately.
The committees identified cost, customization, and near-frontier performance as adoption drivers, and found that the United States lacks a competitive open-weight alternative at the price and performance the market demands.
The report compares published API prices for one million input tokens and one million output tokens (small pieces of words and characters used to measure model usage):
Model Developer Approximate Cost
Qwen3-235B-A22B-Instruct Alibaba $1.15
DeepSeek-R1 DeepSeek $2.74
Kimi K2.6 Moonshot AI $4.95
Claude Opus 4.8 Anthropic $30
GPT-5.5 OpenAI $35
Those figures do not adjust for differences in capability, caching, batch discounts, promotions, or self-hosting costs. They illustrate the price pressure behind adoption, while open weights also let companies fine-tune models on their own data and deploy them on infrastructure they control.
The security concern is separate from hosting location. Running models on U.S. infrastructure reduces direct data exposure, but the report says that does not change how the underlying model was trained or how it may behave under conditions ordinary evaluations miss.
Proposed Restrictions and Replacement Risks
The report’s 12 recommendations do not themselves enact restrictions. They would build on a Department of Homeland Security order excluding DeepSeek products and services from federal civilian procurement and requiring their removal from agency systems, including third-party products built on DeepSeek. The chairmen recommend extending those restrictions to other Chinese-developed models and their derivatives across federal information systems and federal contract work, with narrow waivers for research, testing, and defensive security. They also recommend barring those models and derivatives from narrowly defined high-risk critical-infrastructure functions, with removal schedules tied to trusted alternatives and temporary exceptions where no trusted substitute exists.
To make the rules enforceable, a model-provenance standard led by the National Institute of Standards and Technology and a machine-readable AI bill of materials would document base models, significant fine-tuning, and distillation sources. The requirement would cover federal contractors and AI used in high-risk critical-infrastructure functions, and would be voluntary elsewhere. A national American open-weight AI initiative would combine public computing resources, shared evaluation infrastructure, voluntary licensing of American frontier-model outputs to qualified open-weight developers for training, and federal anchor demand.
The committees warn that removing Chinese models before trusted open-weight alternatives exist could raise costs and concentrate the market among closed proprietary providers. Anysphere’s exploratory plan therefore illustrates the larger implementation test: restrictions need reliable lineage records to locate embedded models and credible substitutes to avoid trading Chinese open weights for higher costs or closed-model dependence.
AI Trends