China open-sources its AI because giving the models away wins on every aspect, except immediate revenue: distribution, standards, cloud customers and geopolitical reach — at the precise moment US sanctions make selling closed Chinese AI abroad almost impossible. The stack strategy is working. Chinese-origin models hit 46.4% of all tokens routed through OpenRouter by mid-2026, ahead of US models at 35.7%, according to OpenRouter data reported by Forkast in July 2026.[1] Hugging Face data shows Chinese models took roughly 41% of global open-model downloads in 2025, overtaking US models for the first time, and China’s industry ministry (MIIT) now claims the country ranks first worldwide in open-source model downloads.[2][3] Nearly the entire Chinese frontier ships open — DeepSeek under MIT licence, Alibaba’s Qwen under Apache 2.0, Zhipu’s (Z.ai’s) GLM-5.2 under MIT, Moonshot AI’s Kimi K3 with open weights — while nearly the entire US frontier (OpenAI, Anthropic, Google’s flagship tiers) ships closed. That inversion is not charity and it is not ideology. It is at least four overlapping strategic logics, and understanding them is the key to every Chinese-AI headline of 2026 — including the signs since July 2026 that the strategy may be about to change.
Free models sell expensive clouds
The first logic is the oldest play in platform economics: commoditise your complement. Alibaba does not need Qwen to make money; it needs Qwen to sell Alibaba Cloud compute, and Qwen workloads have become a visible driver of Alibaba Cloud’s revenue growth (Constellation Research, May 2025).[4] ByteDance monetises attention, not model APIs, so free Doubao-family models that make its products stickier cost it nothing strategically. For Chinese platforms whose real businesses are cloud, commerce and advertising, the model is a loss-leader that grows the pie they actually own — a dynamic we map in our China AI strategy pillar.
The commoditisation is brutal on price. DeepSeek’s V4 Flash lists at $0.14 per million input tokens against $5.00 for OpenAI’s GPT-5.5 — a 35x gap — and OpenRouter data show Chinese models running 60-90% cheaper than comparable OpenAI and Anthropic models (CIO, July 2026).[5][6][1] Martin Casado, general partner at Andreessen Horowitz, told The Economist in August 2025 there was an “80% chance” that startups pitching a16z with open-source models were using Chinese ones.[7] Track how far the floor keeps falling in our AI model cost tracker.
Open weights are sanctions-proof distribution
The second logic is geopolitical. When Washington blocks your access to top-end Nvidia chips and blacklists your labs — Zhipu AI was added to the US Entity List in January 2025 — you cannot compete on serving the world’s best closed model from Chinese data centres.[8] What you can do is publish the weights and let the world run them on its own hardware. Distribution beats margin: an open model needs no US market access, no export licence and no Chinese server to spread.
Washington has noticed that its own tools do not reach this channel. The US-China Economic and Security Review Commission’s (USCC) 23 March 2026 paper, “Two Loops: How China’s Open AI Strategy Reinforces Its Industrial Dominance”, warned that US export controls “primarily target the digital loop” of training compute and are “not well suited to addressing the physical loop” of deployment-driven data, even as cheap, open Chinese models accelerate global uptake.[9] In other words: chip sanctions constrain how Chinese models are made, not how far they travel. The open release is China’s asymmetric answer to the chip war — the same asymmetry we track in the China vs US AI race.
Every derivative is a vote for a Chinese standard
The third logic is standards export. Alibaba’s Qwen family has spawned more than 100,000 derivative models on Hugging Face, the largest model ecosystem on the platform (USCC, March 2026).[9] Every fine-tune, distillation and sovereign-AI project built on a Chinese base locks that architecture, tokeniser and tooling deeper into the global stack. South Korea learned how deep in January 2026, when Naver’s “sovereign” HyperCLOVA X SEED model was found to incorporate a vision encoder showing 99.5% cosine similarity to Qwen components — inside a $1 billion Korean programme meant to build models from scratch.[10] Across Southeast Asia, national and regional models routinely start from Qwen or DeepSeek checkpoints, a pattern we document in our map of Chinese AI models in Southeast Asia.
Beijing has made the export explicit state policy. China’s “AI+” guideline, issued by the State Council in August 2025, calls for a globally open, open-source technology system and community ecosystem, and even encourages universities to count open-source contributions towards student credit.[11][12] At the World AI Conference in Shanghai on 17 July 2026, Xi Jinping launched the World Artificial Intelligence Cooperation Organization (WAICO) with 29 member countries, declaring AI “should not be a solo performance by a single country” (Al Jazeera).[13] Official data released later that month put cumulative downloads of Chinese open-source models at more than 10 billion.[14] Open weights are China’s offer to the Global South; WAICO is the institution built to receive it.
The trust problem open weights solve
The fourth logic is the quietest and, outside China, the most decisive. Chinese consumer software abroad dies on one objection: your data goes to Beijing. TikTok, WeChat Pay and Temu have all fought that fight. Open weights dissolve it. A bank in Jakarta or a telco in Riyadh can download GLM-5.2 or Qwen, run it entirely on its own servers, and truthfully tell regulators that not one byte touches China. Self-hosting converts a Chinese product into neutral infrastructure — which is precisely why Chinese models win enterprise deployments in markets where Chinese apps cannot get licensed.
US security researchers argue the trust is only partial. A NIST-backed CAISI evaluation found DeepSeek-based agents were 12 times more likely than US frontier models to follow malicious instructions (NIST, September 2025), and a Booz Allen Hamilton report in 2026 found Chinese models generated more vulnerable code for users identifying as US government employees, raising “sleeper agent” concerns (Fox News, June 2026).[15][16] Open weights also mean anyone — including bad actors — can fine-tune away safety guardrails. The rebuttal writes itself, though: weights you can inspect and air-gap are still more auditable than a closed API in someone else’s jurisdiction, whichever flag it flies.
Will China stay open once it leads?
Openness was a challenger’s strategy — Nathan Lambert of Interconnects has argued Chinese labs opened up primarily “to close the gap” with US rivals and capture markets that spend less on software. The crucial test, as Lambert framed it in 2025, is what happens when they stop being challengers. Through 2026, the evidence has been mixed and getting more interesting. Alibaba released Qwen3.6-Max-Preview on 20 April 2026 as API-only — the first closed-weight flagship in Qwen’s history — while keeping smaller Qwen3.6 models openly licensed.[17] Moonshot AI launched Kimi K3, a 2.8-trillion-parameter model, via paid API on 16 July 2026 at $3/$15 per million tokens, and published the weights on schedule on 27 July under its own custom licence rather than the Modified MIT terms of earlier Kimi releases.[18][19][20] Zhipu, by contrast, shipped GLM-5.2 with a 1-million-token context window under a clean MIT licence in June 2026, and used its Hong Kong listing (January 2026) to fund more of the same.[21][22]
The bigger signal came from the state. On 21 July 2026, the Financial Times and Reuters reported that China’s Ministry of Commerce was consulting Alibaba, ByteDance and Zhipu on export controls that could restrict foreign access to advanced Chinese model weights — mirroring the very US playbook China’s openness was designed to route around.[23] On 16 October the ministry published a draft: 23 AI technologies proposed for China’s export-control catalogue, with model weights, checkpoints and source code covered once a model passes certain performance thresholds.[24][25] API use and already-published open weights sit outside it, and the ministry says new open-source releases will be cleared as soon as they’re filed.[25] Nothing is final, and Lambert still expects China to stay broadly open because free weights suppress “disorderly competition” and duplicated training costs at home. But the pattern of 2026 — closed flagships, custom licences, a weights-export consultation four days after Xi preached openness at WAIC and a draft rule three months later — suggests the honest answer: China is open because it is winning the adoption war, and openness will last exactly as long as it keeps winning it. Watch where the frontier actually sits in our US vs China capability tracker.
FAQ: China’s open-source AI strategy
Are Chinese AI models really open source?
Mostly open-weight rather than fully open source: DeepSeek and GLM-5.2 use the permissive MIT licence and Qwen’s open line uses Apache 2.0, but training data and code are rarely released. That is still far more open than the US frontier, where OpenAI, Anthropic and Google keep flagship weights closed as of July 2026.
How do Chinese AI labs make money on free models?
Through the complements: Alibaba sells cloud compute, ByteDance sells advertising, and labs such as Moonshot and Zhipu sell hosted APIs and coding subscriptions on top of their open weights — Kimi K3’s API launched at $3/$15 per million tokens in July 2026.[18][19] Monetisation remains the strategy’s weakest link, which is why Z.ai and MiniMax raised public capital in Hong Kong in January 2026.[22][26]
Is it safe for companies outside China to use Chinese open models?
Self-hosted weights send no data to China, which neutralises the biggest objection. But a NIST-backed study found DeepSeek agents 12 times more likely to follow malicious instructions than US frontier models,[15] so tracing a model’s base-model origin before enterprise deployment is prudent.
How much of global AI usage runs on Chinese models?
By mid-2026, Chinese models accounted for 46.4% of tokens routed via OpenRouter versus 35.7% for US models (Forkast, July 2026), took roughly 41% of Hugging Face downloads in 2025, and — per official Chinese data released in July 2026 — have been downloaded over 10 billion times cumulatively.[1][2][14]
Could China restrict its open models in future?
Yes — in October 2026 China’s commerce and science ministries proposed adding the weights of its most capable models to the export-control catalogue, though already-published open weights and API use fall outside the draft.[24][25] Combined with Alibaba keeping its Qwen3.6-Max flagship closed in April 2026, it is the clearest sign yet that openness is a strategy, not a principle.[17]
Sources & Further Reading
- Chinese AI Models Now Capture Up to 46% of US Enterprise Token Usage — Forkast, 2026-07-10
- State of Open Source on Hugging Face: Spring 2026 — Hugging Face, 2026-03-18
- MIIT minister: Value of China’s AI industry hit 1.2 tln yuan in 2025 — CGTN, 2026-03-05
- Alibaba Cloud Q4 growth strong as Qwen, AI workloads extend reach — Constellation Research, 2025-05-15
- Chinese AI models are gaining ground in the U.S., attracting companies with lower prices — CIO, 2026-07
- DeepSeek V4 Flash API pricing — ZenMux, 2026-08-05
- 80% Of Startups Using Open-source Models Are Using Chinese Models: Andreessen Horowitz Partner — OfficeChai, 2025
- US Blacklists Tencent-Backed AI Startup as Tech Curbs Deepen — Bloomberg, 2025-01-16
- Two Loops: How China’s Open AI Strategy Reinforces Its Industrial Dominance — U.S.-China Economic and Security Review Commission, 2026-03-23
- Korea’s AI sovereignty fails: Naver used Chinese model — Byteiota, 2026-01-14
- China to promote AI Plus initiative (State Council guideline) — State Council of China (gov.cn), 2025-08-27
- The State Council issues the “Opinions on Deepening the Implementation of the ‘Artificial Intelligence+’ Action” — TAIBO, 2025-08-26
- China’s Xi Jinping launches new AI alliance: What is it? — Al Jazeera, 2026-08-18
- China’s open-source AI models power real economy — CGTN, 2026-07-31
- CAISI Evaluation of DeepSeek AI Models Finds Shortcomings and Risks — NIST, 2025-09-30
- Chinese AI models raise ‘sleeper agent’ fears after report finds more vulnerable code for US users — Fox News, 2026
- Qwen 3.6 Max Preview: Alibaba’s closed-model pivot — Digital Applied, 2026-04
- Moonshot releases 2.8-trillion-parameter Kimi K3 — Tom’s Hardware, 2026-07
- Kimi K3: 2.8T Parameters, 104B Active, 1M Context, Benchmarks, Pricing (2026) — Morph, 2026
- moonshotai/Kimi-K3 — Simon Willison’s Weblog, 2026-07-27
- Z.ai’s open-weights GLM-5.2 beats GPT-5.5 on multiple long-horizon coding benchmarks for 1/6th the cost — VentureBeat, 2026-06
- China ‘AI tiger’ Zhipu debuts in Hong Kong IPO — CNBC, 2026-01-08
- China considers tighter export controls on AI models and chips, FT reports — The Standard (Reuters), 2026-07-21
- China Considers Export Controls on AI Models, Data and Chips — TechRepublic, 2026-10
- China Drafted Export Controls on AI Model Weights. Open Source Gets a Free Pass. — Let’s Data Science, 2026-10-16
- MiniMax’s Stock Doubles in Value as Chinese AI Startup Debuts in Hong Kong — Yicai Global, 2026-01-09