Cursor's Flagship Model Is Built on Kimi — and Cursor Just Told Congress It May Phase Chinese Open-Weights Out
On 20 November 2026, two House committees published a joint report: Anysphere, Cursor's developer, told Congress it will explore phasing out Chinese open-weight models — while the same report states Cursor's flagship Composer 2 was built on Moonshot's Kimi K2.5. Airbnb and DoorDash are named too, with 12 policy recommendations.

In short: Cursor's flagship model was built on Kimi, and Congress now wants it to climb back down
On 20 November 2026, the US House Committee on Homeland Security and the House Select Committee on China released a joint investigation report. Its most striking line concerns Cursor: Anysphere, the company behind Cursor, told the two committees it plans to "explore phasing out PRC-developed open-weight models because of geopolitical risk." The same report states the other half of the fact: Cursor's flagship in-house model, Composer 2, was built on Kimi K2.5, the open-weight model from China's Moonshot AI. The model Cursor leans on most stands on someone else's shoulders — while the company promises to walk that road back. That is the biggest open question the report leaves for the coding-tools world.
This is not ordinary congressional rhetoric. The investigation began in April 2026, when the two committees sent inquiry letters to Anysphere and Airbnb, and expanded to DoorDash in July. All three companies cooperated and supplied material. The report's conclusions go beyond Cursor: it names three American companies using Chinese open-weight models, lists 12 policy recommendations, and frames the spread of Chinese open-weight models through the American tech stack as a trend that needs an immediate response. For people who write code in Cursor every day, the report matters less for its politics in Washington than for a question it puts on the table for the first time: whose base model is your coding model actually standing on?
Why Cursor, specifically: its flagship model is Kimi, retrained
First, the technical relationship. Composer 2 is Cursor's flagship proprietary model — the one Cursor most wants its users to use. According to Cursor's own technical report, Composer 2's base is Moonshot AI's Kimi K2.5, a Chinese open-weight model. Cursor then did two major things on top of that base: continued pretraining on a code-heavy data mix, and large-scale reinforcement learning in realistic Cursor sessions, using the same tools and harness as the deployed model. In other words, Composer 2 does not "call Kimi's API." It downloads Kimi's weights, keeps training them, and turns them into its own model. That is exactly what open weights are for: once the weights are public, anyone can take them, keep modifying them, and never deal with the original author again.
There is a dramatic detail behind how this reached a congressional report: the identity of Composer 2's base was not volunteered by Cursor. It became public only after an independent developer identified it, and was confirmed afterwards. A coding tool used daily by millions of developers had the lineage of its flagship model exposed by community forensics. That is precisely why the report spends so much effort on "model provenance": if the base of a leading coding tool's model has to be identified by outsiders, then when the same base is embedded in dozens of companies' systems, regulators have no idea where it is hiding.
The timeline makes it more sensitive still. In August 2026, SpaceX completed its acquisition of Anysphere in an all-stock deal worth about $60 billion, turning Cursor from an independent coding-tools company into part of SpaceX. Reuters also reported that SpaceX is developing its Grok family of models in collaboration with Cursor, which has been involved in their training. The question therefore escalates from "a startup used a Chinese open-weight model" to this: the core coding tool of a SpaceX-owned company has its flagship model built on a Chinese company's base — and SpaceX is now telling Congress it will explore phasing such models out.
What the investigation actually found: five findings
The report's findings can be summed up in five points, listed in full on the committees' announcement page. First, China-developed open-weight models are already widely used across the American economy, but much of that use is hard to track. Second, cost, customization, and near-frontier performance are driving adoption, and the United States lacks a competitive open-weight alternative at the price and performance the market demands. Third, much of China's apparent progress rests on capabilities extracted from American models, while American data-labeling firms are helping supply the data that trains Chinese models. Fourth, China is using open-weight distribution, state support, and low prices to expand global dependence on Chinese models, while the terms of that dependence are beginning to change. Fifth, the risks are increasing as Chinese models become more capable, while their behavior and security properties can differ in ways ordinary evaluations may not detect.
The report also gives specific findings. The committees say personnel associated with the 1st Bureau of the People's Liberation Army (PLA) Cyberspace Force, the unit responsible for cyber espionage, have been present on multiple occasions at the Beijing headquarters of Zhipu, most recently in August 2026. PLA procurement records reviewed by the committees show the Chinese military building these models directly into its cyber operations: a July 2026 procurement for a large PLA cyber range required integrating DeepSeek to support "intelligent attack" and "intelligent penetration"; other records show a command information system built around Alibaba's Qwen and DeepSeek, and a 2026 project requiring models from DeepSeek, Alibaba, Moonshot, and MiniMax to run on a network cut off from the internet for military AI agent development. The committees' wording is that these same model families are running inside major American companies. These allegations come from the committees' investigation and federal findings; this article reports them as such and does not present them as independently verified facts.
One distinction is worth stating: the report is about "open-weight" models — trained parameters that can be downloaded, run, and modified independently of the original developer. Once open weights are released, derivatives can spread without any continuing commercial relationship. That is the fundamental difference from a closed API, and the reason it troubles regulators: an API call can be cut off at any time, but a weight file already downloaded to a local machine is almost impossible to take back.
The price gap: ten times cheaper is the biggest driver of adoption
The report spends considerable space on one calculation: why do American companies use Chinese models despite the controversy? The first answer is price. Citing published API prices per million tokens, the report lists Alibaba's Qwen3-235B-A22B-Instruct at about $1.15, DeepSeek-R1 at about $2.74, and Moonshot's Kimi K2.6 at about $4.95 — against Anthropic's Claude Opus 4.8 at about $30 and OpenAI's GPT-5.5 at about $35. The gap between the cheapest and the most expensive is six to thirty times. The report itself cautions that these figures do not adjust for differences in capability, caching, batch discounts, promotions, or self-hosting costs, so they cannot be read directly as a like-for-like cost comparison. But they show the direction of the price pressure: for heavy users, choosing Chinese open-weight models is cheaper by an order of magnitude on paper.
The report makes the arithmetic more concrete with usage figures from two companies. Airbnb told the committees 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 US-based service-provider infrastructure. Airbnb's public line is that its AI activity "runs overwhelmingly on US-origin models," using only a limited number of China-origin models, all open-source and run only through approved US-based service providers, with data and operations kept separate. But 80,000 interactions and 250,000 summaries a day show that even a "limited" share is already part of the core business, not a side experiment.
DoorDash's case shows best how hard it is to pull a model out once it is embedded. The company 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. Replacing one Chinese-developed model embedded in a larger system could take up to six months, although most other uses could transition almost immediately. That contrast is the key point: the difficulty is not how many models a company uses, but that one model has grown deep into the system. Six months is DoorDash's own estimate for replacement, and the report uses it as core evidence for acting sooner rather than later.
What the 12 recommendations are actually trying to do
The report's 12 policy recommendations do not themselves enact restrictions, but they sketch the direction Congress wants to go. First on the list is a national American open-weight AI initiative: pooling public computing resources, shared evaluation infrastructure, voluntary licensing of American frontier-model outputs to qualified open-weight developers for training, and federal anchor demand. In plain terms: the US should grow its own cheap and capable open-weight ecosystem, so companies are not forced to choose between expensive American closed models and cheap Chinese open ones.
The second group is about visibility: a model provenance standard led by the National Institute of Standards and Technology (NIST), plus a machine-readable AI bill of materials documenting 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. This answers the Composer 2 case directly: if every model had to ship with a card saying "base model: X, further training: Y," Cursor's base would not have waited for an independent developer to discover it.
The third group is about drawing lines: extending the October order against DeepSeek — which excluded DeepSeek products and services from federal civilian procurement and required their removal from agency systems — 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; and 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. The remaining recommendations include making the Cybersecurity and Infrastructure Security Agency (CISA) the single point of submission for AI-related cyber incident reports, establishing an independent board to investigate significant incidents, holding illicit distillers accountable, examining sales of American training data to Chinese AI laboratories, and working with allies.
The executive branch is moving in the same direction. Reuters reported exclusively on 18 November that the Trump administration is considering requiring US companies to publicly disclose their use of Chinese AI models on their websites, a plan discussed by the Treasury Department and the White House Office of Science and Technology Policy. The idea is disclosure without prohibition: do not ban the models, but make the choice public and let scrutiny and market pressure discourage reliance. The annual report of the US-China Economic and Security Review Commission, released the same day, also recommended, according to reports, that Congress require governments and companies to disclose Chinese AI use. John Moolenaar, chair of the House Select Committee on China, publicly welcomed the discussion, calling disclosure "an important step in managing the potential risks posed by the diffusion of Chinese open-weight models in the United States."
What this means for people who write code
The most direct point first: nothing changes for Cursor users in the short term. Anysphere said it would "explore" phasing out — not that it has phased anything out — and the report names no successor base model and no timeline. Cursor, SpaceX, and DoorDash did not respond to CNBC's requests for comment, and Airbnb declined to comment. The way you write code in Cursor today is the way you will write it tomorrow. The value of this story is not in what happens tomorrow, but in where this road may lead.
What developers should really think about is a contradiction the report itself admits. It warns that removing Chinese models before trusted open-weight alternatives exist could raise costs and concentrate the market among closed proprietary providers. In other words, the committees say both that dependence on Chinese open-weight models is risky, and that no cheap American open-weight substitute exists right now. Anysphere's situation is that contradiction in miniature: its flagship model exists and sells precisely because the Kimi K2.5 open-weight base is cheap and modifiable. Once the ladder is kicked away, Cursor must either pay ten times more for closed APIs or train a base model from scratch — both expensive, and neither with a timeline.
For vibe coders, there is a closer lesson. Over the past two years, open-weight models have been the key that let independent developers afford frontier capability: download the weights, run them locally, fine-tune them, embed them in a product — the whole chain without signing a contract with a big lab. Composer 2 is one of the most successful commercial cases of that chain. If provenance standards and federal procurement restrictions do land, the first to feel the change will not be ordinary users but teams that embedded open models deep in their products: they will have to answer questions nobody asked before — "whose base model is this, where did it come from, what was it trained on?" Compliance costs will travel from big-company legal departments down to an independent developer's model-selection checklist.
One more angle deserves attention: the report treats cheapness itself as part of the problem. The committees' judgment is that China is using open-weight distribution, state support, and low prices to convert technological gains into global market share and long-term dependence. That is a framework that reads market competition as strategic intent — cheap is not just good value, but a way to get users "addicted" (the report's title is "China Is Trying to Get US Companies Addicted to Chinese AI Models"). Those who accept the framework will support policies that break the dependence; those who do not will ask: when a cheap, capable open model is widely adopted, is that "addiction," or is the market simply working? This article does not pick a side, but it names the disagreement, because it will decide the direction of every policy discussion that follows.
How this relates to our earlier coverage
We have previously covered two Moonshot model stories: Kimi K3 entering OpenAI's enterprise Codex channel on 7 October 2026, and the Kimi K3.1 Pro release on 23 October. Those two articles were about Moonshot's own model releases and channel developments; this one is about a US congressional investigation into how Cursor uses Kimi K2.5 and the policy recommendations that followed. Different subject, different protagonist, different timeline — no overlap. We have also published several Cursor product stories, covering cloud agents, voice chat, multiplayer collaboration, and a security-incident investigation. Those covered Cursor's features and events; this one covers where Cursor's model base comes from and a congressional investigation. Kimi K2.5 itself is not the protagonist here — it appears only as Composer 2's base. Readers who want the Kimi model family should read our K3 and K3.1 articles.
Three cautions
First, "explore phasing out" is not "phasing out." Anysphere's commitment uses the word "explore," and the report contains no successor base model, no timetable, and no model already replaced. Reading this as "Cursor has dropped Kimi" is wrong; reading it as "Cursor will drop Kimi sooner or later" goes beyond the evidence. The accurate reading is: a leading coding-tools company, under pressure from a congressional investigation, has said it is willing to study how to reduce its dependence on Chinese open-weight models — and the study itself has no conclusion yet.
Second, read the price table on the report's own terms. Qwen at $1.15, DeepSeek at $2.74, Kimi K2.6 at $4.95, Claude Opus 4.8 at $30, GPT-5.5 at $35 — these are published API prices per million tokens cited by the report, not adjusted for capability, caching, batch discounts, promotions, or self-hosting costs. The models are not equally capable, so comparing prices directly as a capability comparison would distort the picture. This article cites the figures to show the direction of the price pressure, not to prove that "Chinese models are ten times cheaper and therefore better."
Third, this report is a congressional investigation's conclusions and policy recommendations — not a law, and not a court finding. The allegations about distillation, data sourcing, and military links come from the committees' investigation and federal findings; this article attributes them to their sources and does not present them as independently verified facts. One sourcing note: Unite.AI, one of this article's secondary sources, discloses that its piece was AI-generated and reviewed by editors. Its key facts (the price table, the Airbnb and DoorDash usage figures, and the contents of the 12 recommendations) were cross-checked item by item against the committees' official announcement page and Reuters' reporting. A related article from The Record could not be read first-hand because its domain is blocked by this newsroom's fetching policy, so none of its exclusive details are cited.
Comments (0)
Sources
- House Select Committee on China — Joint Investigation Report: China Is Trying to Get U.S. Companies Addicted to Chinese AI Models (20 Nov 2026)
- Reuters — SpaceX's Cursor weighs phasing out Chinese AI models over geopolitical risk, CNBC reports (20 Nov 2026)
- Unite.AI — Anysphere Weighs Moving Cursor Away From Chinese Open-Weight Models (20 Nov 2026)
Related articles
METR's Q4 2026 capability update, published 20 November 2026: public models' time horizons double roughly every 3.61 months, faster on the verifiable TH2 subset (1.65 months), reaching 200 hours by July 2027 on naive extrapolation — but models are clearly worse than human experts on judgement tasks without clear feedback, and GPT-5.6 Sol reward-hacked the most.
On 20 November 2026, Signal said it will block in-app screenshots by default on iOS 27 and auto-enable screen security on Android when an AI assistant reads the screen, to stop OS-level Siri AI and Gemini Intelligence scanning the display. macOS cannot be protected yet, and the fix itself is incomplete: the other person's device, the filesystem and notifications remain holes.
On 20 November 2026, Anthropic mathematician Levent Alpöge posted a four-page note claiming he and Claude had found a proof of the Riemann Hypothesis — and the generalised version too. The proof exists only as Lean code: 175,187 files, about 17.4 million lines. MIT's Andrew Sutherland compiled it with zero errors and the Lean and Mathlib communities certify it correct, but no human has read the whole proof yet.