Companies using Chinese AI models
US companies and agencies that use, build on, or test AI models from Chinese developers such as Alibaba and Moonshot AI, with the date and the source for each.
Updated · ledger as of Nov 18, 2026
US companies are building products on AI models from Chinese developers, mostly Alibaba's Qwen, Moonshot AI's Kimi, and DeepSeek. Most of these models are open-weight, and companies cite their low cost. Two House committees have written to Airbnb, Cursor maker Anysphere, and DoorDash about those uses, and on Nov. 18, 2026, Reuters reported that the Treasury Department and the White House are discussing a plan that would require companies to disclose them on their websites. This page lists every use we can source, with the model, what it does, the date, and the source.
Who uses Chinese AI models
Newest first. A row is a disclosed or reported use, not a finding of wrongdoing. Most of these models are open-weight, so the companies can run them on US servers.
| Company | Model | What it does | Status | Date | Source |
|---|---|---|---|---|---|
| Apple (AAPL) | QwenAlibaba | iPhone users in mainland China will be able to use Qwen through Siri, Writing Tools, and Visual Intelligence when iOS 27.3 comes out in December. Testing began in the iOS developer beta on Nov. 4. The arrangement covers China only, and the rest of Apple Intelligence runs on Apple's own models. | PlannedApple announced it | Oct 31, 2026 | Our report Apple newsroom (Chinese) |
| Federal Register websiteRun by the National Archives | QwenAlibaba | A search tool for browsing public comments on proposed federal regulations, offered alongside other search options.Since then: Taken down on Sept. 16, 2026, around the time social media posts pointed it out, Reuters reported. It was unclear when the tool was deployed. | StoppedReported by Reuters | Sep 17, 2026 | Reuters (via Yahoo) |
| HarveyLegal AI company | Kimi K3Moonshot AI | Harvey Tenet, Harvey's first in-house legal model, is a Kimi K3 base that Harvey post-trained with Fireworks for long-horizon legal work. Harvey described it as research it will fold into its product over time. | Built onHarvey said so | Aug 20, 2026 | Harvey blog |
| GitHub Copilot (MSFT)GitHub is owned by Microsoft | Kimi K3Moonshot AI | Offered as a model users can pick in GitHub Copilot on the Pro, Pro+, Max, Business, and Enterprise plans. GitHub hosts the model on Fireworks AI. It is off by default for Business and Enterprise customers. | In useGitHub announced it | Aug 6, 2026 | GitHub changelog |
| Replit | Kimi K3Moonshot AI | The first model in Replit's model choice feature, which lets users pick the model behind Replit's coding agent. | In useCEO Amjad Masad said so | Jul 28, 2026 | Amjad Masad on X |
| Microsoft (MSFT) | Kimi K3Moonshot AI | Copilot engineers planned to evaluate whether Kimi K3 could power features that run on OpenAI and Anthropic systems, a shift that could cut Microsoft's AI inference costs by as much as $600 million, The Information reported. Microsoft was also adding Kimi K3 to its Azure cloud.Since then: No report since has said whether Copilot adopted the model. Microsoft hasn't publicly confirmed the report. | TestingReported by The Information | Jul 20, 2026 | Our report The Information CryptoBriefing |
| Hugging Face | GLM 5.2Z.ai | Forensic analysis of a July 2026 intrusion into Hugging Face's infrastructure. AI providers' safety guardrails blocked requests that included real attack commands and exploit payloads, so Hugging Face ran the open-weight GLM 5.2 on its own servers. | In useHugging Face said so | Jul 16, 2026 | Our report Hugging Face disclosure |
| CognitionMaker of Devin and Windsurf | Kimi K2.7, Kimi K3Moonshot AI | Cognition's SWE-1.7 coding model was trained from a Kimi K2.7 base, and SWE-2, released Sept. 10, 2026, is post-trained from Kimi K3. Both run in Devin, Cognition's coding agent. | Built onCognition said so | Jul 8, 2026 | Cognition: SWE-1.7 Cognition: SWE-2 |
| DoorDash (DASH) | Kimi K2.6Moonshot AI | Handles lower-level work in DoorDash's internal AI code reviewer, while Anthropic's Fable does the hardest work. | In useCo-founder Andy Fang said so | Jul 6, 2026 | Our report Andy Fang on X House committees' letter |
| Coinbase (COIN) | GLM 5.2, Kimi 2.7Z.ai and Moonshot AI | Coinbase is experimenting with making these open-weight models the defaults in its internal LLM gateway, while engineers can still choose other models. CEO Brian Armstrong said the changes, along with routing and caching, cut Coinbase's AI spending nearly in half. | TestingCEO Brian Armstrong said so | Jun 26, 2026 | Brian Armstrong on X Fortune |
| LindyAI agent startup | DeepSeek V4DeepSeek | All of Lindy's traffic, replacing Anthropic's models. CEO Flo Crivello said the switch saves Lindy millions of dollars. | In useCEO Flo Crivello said so | Jun 4, 2026 | Flo Crivello on X |
| Uber (UBER) | QwenAlibaba | The backbone of the embedding model behind Uber Eats search, fine-tuned on Uber's own data. Uber said the single model handles content embedding for all Uber Eats verticals and markets. | In useUber Engineering said so | Apr 14, 2026 | Uber Engineering blog |
| CursorMade by Anysphere, which SpaceX bought on Aug. 14, 2026 | Kimi K2.5Moonshot AI | Cursor's Composer 2 coding model, launched in March 2026, was trained starting from Kimi K2.5, through continued pretraining and reinforcement learning. Cursor reaches the model through Fireworks AI under an authorized commercial partnership, Moonshot said. | Built onCursor and Moonshot said so | Mar 20, 2026 | Cursor technical report Moonshot AI on X |
| Meta Platforms (META) | QwenAlibaba | Meta's TBD Lab was distilling from Qwen, along with Google's Gemma and OpenAI's gpt-oss, to help train its Avocado model, Bloomberg reported. | Used in trainingReported by Bloomberg | Dec 10, 2025 | Bloomberg Sherwood News |
| Nous Research | Seed-OSS-36BByteDance | Hermes 4.3 36B, Nous Research's open-weight model, is based on ByteDance Seed's Seed-OSS-36B-Base. | Built onNous Research said so | Dec 1, 2025 | Nous Research blog |
| Airbnb (ABNB) | QwenAlibaba | Powers Airbnb's AI customer service agent, which is built on 13 AI models, including ones from OpenAI and Alibaba.Since then: On Nov. 18, 2026, Airbnb told Reuters that its AI activity runs overwhelmingly on US-origin models and that the few China-origin models it uses are open-source and run only through approved US-based service providers. | In useCEO Brian Chesky said so | Oct 21, 2025 | Our report Fortune and Bloomberg Reuters, Nov. 18, 2026 |
| Tesla (TSLA) | Doubao, DeepSeek ChatByteDance and DeepSeek | The in-car voice assistant in Tesla's cars in China. Doubao handles most voice commands and DeepSeek Chat handles AI conversation, according to a Tesla China document Bloomberg reviewed. The arrangement covers China only.Since then: Tesla began bringing Doubao to its cars in China in July 2026, CnEVPost reported. | In useReported by Bloomberg | Aug 22, 2025 | Bloomberg CnEVPost, July 31, 2026 |
| Department of EnergyTwo unnamed national labs | DeepSeekDeepSeek | Two DOE national labs ran controlled, documented studies comparing DeepSeek with their own alternatives. DOE deputy CIO Bridget Carper said the labs found some positives the agency may approve, though DeepSeek isn't allowed across DOE. | TestingDOE's deputy CIO said so | Aug 19, 2025 | FedScoop |
| Nvidia (NVDA) | Qwen2.5, Qwen3Alibaba | Several Nvidia open models are built on Qwen: the OpenReasoning-Nemotron reasoning models (the 32B version is a derivative of Qwen2.5-32B) and the Cosmos Reason2 robotics model (from Qwen3-VL-8B). Nvidia also fine-tuned Qwen3-235B into a reward model used to train Nemotron 3 Nano. | Built onNvidia's model cards say so | Jul 15, 2025 | OpenReasoning-Nemotron card Nemotron reward model card |
| Snowflake (SNOW) | Qwen2.5-Coder-7BAlibaba | Arctic-Text2SQL-R1-7B, Snowflake's open model for turning questions into SQL, is fine-tuned from Qwen2.5-Coder-7B-Instruct. | Built onSnowflake's model card says so | May 22, 2025 | Snowflake model card |
| Microsoft (MSFT) | DeepSeek-R1DeepSeek | MAI-DS-R1, a version of DeepSeek-R1 that the Microsoft AI team post-trained to respond on topics R1 blocked and to improve its risk profile, released as an open model. | Built onMicrosoft's model card says so | Apr 16, 2025 | Microsoft model card |
| Perplexity | DeepSeek-R1DeepSeek | R1 1776, a version of DeepSeek-R1 that Perplexity post-trained to remove Chinese Communist Party censorship, released as open weights and offered in Perplexity's API.Since then: Removed from Perplexity's API on Aug. 1, 2025. | StoppedPerplexity said so | Feb 18, 2025 | Perplexity model card (archived) Perplexity API changelog |
| Department of Defense | DeepSeekDeepSeek | Defense Department employees connected work computers to DeepSeek's chatbot for at least two days before the Defense Information Systems Agency moved to block it, Bloomberg reported. Military personnel had downloaded an earlier DeepSeek release onto workstations in fall 2024.Since then: Thousands of Defense Department personnel were still using DeepSeek through Ask Sage, an authorized platform that doesn't connect to Chinese servers, Ask Sage's founder told Bloomberg. | StoppedReported by Bloomberg | Jan 30, 2025 | Bloomberg Bloomberg (via The Star) |
By developer
| Developer | Users | Who |
|---|---|---|
| Moonshot AI | 8 | Harvey, GitHub Copilot, Replit, Microsoft, Cognition, DoorDash, Coinbase, Cursor |
| Alibaba | 7 | Apple, Federal Register website, Uber, Meta Platforms, Airbnb, Nvidia, Snowflake |
| DeepSeek | 6 | Lindy, Tesla, Department of Energy, Microsoft, Perplexity, Department of Defense |
| Z.ai | 2 | Hugging Face, Coinbase |
| ByteDance | 2 | Nous Research, Tesla |
Investigations and disclosure proposals
Congressional letters, federal orders, and proposals on the use of Chinese AI models, newest first.
| Date | What happened | Source |
|---|---|---|
| Nov 18, 2026 | The Treasury Department and the White House Office of Science and Technology Policy have discussed requiring companies such as Airbnb to disclose their use of Chinese AI models on their websites, Reuters reported. It wasn't clear whether the idea was gaining traction or would go through a formal rulemaking. | Our report Reuters |
| Oct 20, 2026 | The Department of Homeland Security published an order requiring federal civilian agencies to stop buying DeepSeek products and to remove DeepSeek products, and third-party products and services developed using DeepSeek, from their IT systems. The signed order lists an original release date of July 24, 2026. | Our report Federal Register |
| Jul 31, 2026 | House Homeland Security Committee Chairman Andrew Garbarino and House China select committee Chairman John Moolenaar wrote to DoorDash CEO Tony Xu about the company's use of Kimi K2.6, asking for documents by Aug. 14 and a briefing by Aug. 21. | China select committee |
| Apr 29, 2026 | Garbarino and Moolenaar announced a joint investigation into the national security risks of Chinese AI models, with letters to Airbnb, about its use of Qwen, and to Cursor maker Anysphere, about Composer 2. | China select committee Letter to Airbnb (PDF) |
How this ledger is kept
This page lists uses of AI models made by Chinese developers, such as Alibaba's Qwen, Moonshot AI's Kimi, and DeepSeek, by US companies and US government agencies. Many of these models are open-weight, which means anyone can download and adapt them, so a company can run one on its own servers or through a US cloud provider.
Each row links to the record we read: the company's own statement or post, an executive's on-the-record remarks, or a report from a major outlet. "Confirmed" means the company or one of its executives said so publicly. "Reported" means an outlet said so and the company didn't confirm it on the record. The date is when the use was first disclosed or reported, not when it began. "About" marks a date we couldn't pin down to the day.
"In use" means the model runs in a product or internal system. "Built on" means the company trained its own model starting from a Chinese one. "Testing" means the company was evaluating the model, which doesn't mean it adopted it. "Planned" means the company has announced a launch that hasn't happened yet. Cloud catalogs that only list Chinese models for developers to rent, such as Microsoft's Azure AI Foundry, aren't counted as a use.
Common questions
- Does Apple use Alibaba's Qwen?
- Yes, in China. Apple said on Oct. 31, 2026, that iPhone users in mainland China will be able to use Alibaba's Qwen through Siri, Writing Tools, and Visual Intelligence starting with iOS 27.3 in December. The rest of Apple Intelligence runs on Apple's own models, and the arrangement doesn't apply outside China.
- Did the US government use a Chinese AI model?
- Yes, briefly. The Federal Register website, which the National Archives runs, offered a search tool built on Alibaba's Qwen for browsing public comments on proposed regulations, Reuters reported on Sept. 17, 2026. The tool was taken down on Sept. 16, around the time social media posts pointed it out.
- Is Harvey's legal AI built on Kimi?
- Harvey Tenet, the legal AI company's first in-house model, is. Harvey said on Aug. 20, 2026, that Tenet is a Moonshot AI Kimi K3 base that it post-trained with Fireworks for legal work, and that it will bring capabilities from that research into its product over time.
- Does GitHub Copilot use Kimi K3?
- Yes. GitHub, which Microsoft owns, made Moonshot AI's Kimi K3 available in GitHub Copilot on Aug. 6, 2026, for the Pro, Pro+, Max, Business, and Enterprise plans. GitHub hosts the model on Fireworks AI, and it is off by default for Business and Enterprise customers.
- Does Replit use Kimi K3?
- Yes. Replit CEO Amjad Masad said on July 28, 2026, that Replit was launching model choice, starting with Moonshot AI's Kimi K3.
- Does Microsoft use Kimi K3?
- Microsoft was preparing to test Moonshot AI's Kimi K3 for Copilot features that run on OpenAI and Anthropic systems, The Information reported on July 20, 2026. No report since has said whether Copilot adopted it. Separately, GitHub, which Microsoft owns, offers Kimi K3 in GitHub Copilot.
- Did Hugging Face use a Chinese AI model?
- Yes. Hugging Face said on July 16, 2026, that it ran the forensic analysis of an intrusion into its infrastructure on Z.ai's open-weight GLM 5.2, on its own servers, after hosted models blocked requests that included real attack data.
- Is Cognition's Devin built on Kimi?
- Cognition's own coding models are. Cognition said SWE-1.7, released July 8, 2026, was trained from Moonshot AI's Kimi K2.7, and SWE-2, released Sept. 10, 2026, is post-trained from Kimi K3. Both run in Devin, Cognition's coding agent.
- Does DoorDash use Kimi?
- Yes. DoorDash co-founder Andy Fang posted on X on July 6, 2026, that DoorDash's internal AI code reviewer sends lower-level work to Moonshot AI's Kimi K2.6 and the hardest work to Anthropic's Fable. Two House committees wrote to DoorDash CEO Tony Xu about it on July 31, 2026.
- Does Coinbase use Chinese AI models?
- Coinbase is experimenting with them. CEO Brian Armstrong posted on June 26, 2026, that Coinbase is trying Z.ai's GLM 5.2 and Moonshot AI's Kimi 2.7 as the default models in its internal LLM gateway, and that its changes cut AI spending nearly in half.
- Which company switched from Anthropic to DeepSeek?
- Lindy, an AI agent startup. CEO Flo Crivello posted on June 4, 2026, that Lindy had switched 100% of its traffic to DeepSeek V4, dropping Anthropic's models, and that the move saves it millions of dollars.
- Does Uber use Alibaba's Qwen?
- Yes, for Uber Eats search. Uber's engineering blog said on April 14, 2026, that a Qwen model is the backbone of the embedding model behind delivery search, and that it supports all Uber Eats verticals and markets.
- Is Cursor built on Kimi?
- Cursor's Composer 2 coding model is. Cursor's technical report says Composer 2 was trained starting from Moonshot AI's open-weight Kimi K2.5, and Moonshot confirmed it on March 20, 2026. Cursor's maker, Anysphere, has been owned by SpaceX since Aug. 14, 2026.
- Did Meta use Qwen to train its AI?
- Bloomberg reported on Dec. 10, 2025, that Meta's TBD Lab was distilling from Alibaba's Qwen, Google's Gemma, and OpenAI's gpt-oss to help train its Avocado model.
- Does Airbnb use Alibaba's Qwen?
- Yes. Airbnb CEO Brian Chesky said in October 2025 that the company was relying a lot on Alibaba's Qwen for its AI customer service agent, and Reuters reported on Nov. 18, 2026, that Qwen still powers it. Airbnb told Reuters that its AI runs overwhelmingly on US-origin models and that the China-origin models it uses are open-source and run through approved US-based providers.
- Does Tesla use DeepSeek or Doubao?
- In China. Bloomberg reported on Aug. 22, 2025, that Tesla planned in-car voice functions in China powered by ByteDance's Doubao and DeepSeek, citing a Tesla document, and CnEVPost reported in July 2026 that Tesla had started bringing Doubao to its cars there.
- Are Nvidia's Nemotron models built on Qwen?
- Some are. Nvidia's model card says OpenReasoning-Nemotron-32B is a derivative of Alibaba's Qwen2.5-32B, and Nvidia's Cosmos Reason2 robotics model is post-trained from Qwen3-VL-8B. Nvidia also built a reward model from Qwen3-235B for training Nemotron 3 Nano.
- Did the Pentagon use DeepSeek?
- Defense Department employees used DeepSeek's chatbot on work computers for at least two days before the Pentagon moved to block it, Bloomberg reported on Jan. 30, 2025. Ask Sage's founder told Bloomberg that thousands of personnel were using DeepSeek through his authorized platform, which doesn't connect to Chinese servers.
- Which Chinese AI models do US companies use?
- The two that show up most on this page are Alibaba's Qwen, used by Airbnb and, in China, Apple, and Moonshot AI's Kimi, used by Cursor, DoorDash, and GitHub Copilot. Chinese models were the three most popular by token usage in October 2026 on OpenRouter, an AI marketplace that tracks usage, Reuters reported.
- Will US companies have to disclose their use of Chinese AI?
- Not yet. Reuters reported on Nov. 18, 2026, that the Treasury Department and the White House Office of Science and Technology Policy have discussed a plan that would require companies such as Airbnb to disclose it on their websites. It wasn't clear whether the idea would go through a formal rulemaking.
- How often is this page updated?
- A company or agency gets a row when it says publicly that it uses, builds on, tests, or plans to offer a model from a Chinese developer, or when a major outlet reports that it does. Rows are added as new disclosures, congressional letters, and reports come out, and existing rows are updated when a company changes course.
Use this data
Our figures are free to reuse in articles, newsletters, and research with attribution and a link back to this page. Current table as a file: download CSV.
Suggested citation: AIStockWire, “Companies using Chinese AI models,” aistockwire.com/companies-using-chinese-ai-models, data as of Nov 18, 2026. Questions about the data: support@aistockwire.com.
Related: Companies blocking AI agents, Data center moratorium tracker, China AI stocks. Not investment advice.