Europe’s most downloaded open model pulls about 625,000 downloads a month. A single Qwen model pulls twenty-five million.
Chinese labs are not winning the open-weight race on scale alone; they are winning it on the thing Europe keeps promising and rarely delivers, which is a licence you can actually build on.
This is the full map of who builds Chinese large language models, who owns them, what their licences permit, and which ones anyone outside China is running. It is the companion piece to our European LLM map, built the same way and on the same day’s data.
Why This Map Follows the Listing
The European map follows money that runs out. Most European projects end when a grant finishes, so we tracked who was paying and for how long.
China needs a different organising principle, because the money did not run out. It went public.
Z.ai, formerly Zhipu, listed in Hong Kong on 8 January 2026 seeking $560m. MiniMax listed the next day seeking up to $539m and doubled on debut. Moonshot is preparing a Hong Kong listing on a round valuing it at just over $30bn, and DeepSeek is targeting Shanghai’s STAR market as early as the second quarter of 2027 at a reported $71bn. Z.ai and MiniMax beat OpenAI and Anthropic to market, with Moonshot and DeepSeek lining up behind them.
This map follows ownership, and what the public markets now expect from each lab. That is a harder discipline than a grant, and it is the reason the release cadence has not slowed.
Every Active Project in One Table
Licences are the column that matters most here, and the one most coverage gets wrong in both directions. Several Chinese flagships ship under MIT or Apache 2.0, which is more permissive than most of what Europe calls open. One major family excludes the European Union outright.
| Project | Backing | Latest model | Licence | Money behind it |
|---|---|---|---|---|
| Alibaba / Qwen | Alibaba Cloud | Qwen3.8-Max / Qwen3.8-2.4T-A95B (2.4T, 95B active); Qwen3.8-27B | Split: Apache 2.0 on 27B; custom Qwen3.8-Max License on the 2.4T checkpoint | Alibaba Cloud balance sheet |
| DeepSeek | High-Flyer; founder Liang Wenfeng | V4-Flash (284B total, 13B active); V4-Pro | MIT | $7.4bn first external round, June 2026, at over $50bn. STAR market IPO targeted for Q2 2027 at a reported $71bn |
| Moonshot AI / Kimi | Alibaba-backed | Kimi K3 (2.8T total, 104B active) | Kimi K3 License (custom) | $2bn at $20bn in May 2026; Hong Kong listing prepared on a round at just over $30bn |
| Z.ai (Zhipu) | Meituan, Tencent, Ant Group | GLM-5.2 (753B). GLM-5.3 shipped 14 August on the same base; API and Coding Plan were live before the pull, weights were not. | MIT | Hong Kong IPO 8 Jan 2026, sought $560m. H1 2025: about $330m of losses on $27m revenue |
| MiniMax | Alibaba, Abu Dhabi Investment Authority, Mirae Asset | MiniMax-M3 (~428B, 23B active). M2.7 (229B) remains the high-download checkpoint. | “Other”, a custom licence | Hong Kong IPO 9 Jan 2026, sought up to $539m, doubled on debut |
| Tencent | Tencent | Hy3 (295B total, 21B active, 256K context), 6 Jul 2026 | Apache 2.0 for Hy3. The older Hunyuan community licence excludes the EU, UK and South Korea | Tencent balance sheet |
| ByteDance Seed | ByteDance | Seed-OSS-36B (512K context); UI-TARS | Apache 2.0 | ByteDance balance sheet |
| Baidu | Baidu, listed | ERNIE 4.5 family (21B-A3B; 28B-A3B VL) | Apache 2.0 | Listed company |
The Graveyard: The Tigers That Stopped
Two of the original “AI tigers” have stopped shipping frontier models. The weights are still there and people still pull them, which is the tell: the download numbers are respectable and the release dates are old.
| Project | Latest meaningful release | What happened |
|---|---|---|
| 01.AI (Yi) | Yi-1.5 family, 2024 vintage | Top repositories are all 2023 and 2024 models. Yi-6B-Chat still pulls 31,451 a month, on nothing new |
| Baichuan | Baichuan2, 2023; an M3 medical model since | The general-purpose line stopped. Baichuan2-7B-Chat still pulls 33,468 a month while the company works in medicine |
The Licence Question Cuts Both Ways
The most permissive licences in open AI right now are Chinese. DeepSeek ship under MIT. ByteDance’s Seed-OSS and Baidu’s ERNIE 4.5 ship under Apache 2.0. Qwen still does at 27B; the 2.4T Max-class checkpoint does not. It uses a custom Qwen3.8-Max License with attribution and MaaS thresholds.
And the one open model family you cannot legally use in Europe is also Chinese. Tencent’s Hunyuan community licence states, in capitals, that “THIS LICENSE AGREEMENT DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA”, and defines its territory as the world minus those three jurisdictions. An EU company running a Hunyuan model under that licence has no grant of rights at all.
Tencent reversed course. The Hy3 preview in April carried the same exclusion; the full release on 6 July 2026 shipped under Apache 2.0. The lesson for anyone choosing a model is that the licence travels with the release, not with the company, and you have to read it every time.
Two families need reading before deployment for a different reason. Moonshot ships Kimi K3 under its own “Kimi K3 License”, and MiniMax labels M2.7 simply as “other”. Neither is a standard permissive licence, so do not assume either one is open just because you can download the weights.
Who Actually Downloads These Models?
We pulled last-30-day download counts from the Hugging Face API on 25 August 2026, the same method and the same day as the European map, so the two are directly comparable.
| Model | Lab | Downloads, last 30 days |
|---|---|---|
| Qwen3-0.6B | Alibaba | 25,235,903 |
| Qwen3-8B | Alibaba | 15,314,166 |
| Qwen3.5-9B | Alibaba | 13,785,751 |
| DeepSeek-R1 | DeepSeek | 4,431,386 |
| DeepSeek-V4-Flash-0731 | DeepSeek | 3,857,140 |
| Unlimited-OCR | Baidu | 2,932,554 |
| Kimi-K3 | Moonshot | 2,921,257 |
| GLM-OCR | Z.ai | 2,599,280 |
| GLM-5.2 | Z.ai | 2,184,066 |
| MiniMax-M2.7 | MiniMax | 913,819 |
| UI-TARS-1.5-7B | ByteDance | 832,887 |
| Apertus-8B-Instruct (Europe, for contrast) | EPFL / ETH Zurich / CSCS | 624,507 |
| HunyuanOCR | Tencent | 607,356 |
| ERNIE-4.5-VL-28B-A3B | Baidu | 95,917 |
| Hy3-preview | Tencent | 60,850 |
| Seed-OSS-36B-Instruct | ByteDance | 38,278 |
| Baichuan2-7B-Chat | Baichuan, dormant | 33,468 |
| Yi-6B-Chat | 01.AI, dormant | 31,451 |
The line to sit with is the Apertus row. Europe’s most downloaded open model, the flagship of a Swiss federal and academic programme, sits between ByteDance’s interface model and Tencent’s OCR model on a Chinese chart. One Alibaba model released for edge devices out-pulls it forty to one.
How Reliable Are These Download Numbers?
- Downloads are not usage. API traffic does not appear, and every one of these labs runs a commercial API alongside the weights.
- Small models dominate. Qwen3-0.6B leads this table because it fits on a phone, not because it is the best model in it. The same pattern runs through the European map.
- OCR and vision models inflate the top. Four of the twelve largest Chinese numbers here are document or interface models pulled by pipelines, not chat models chosen by people.
- We take one lab’s headline repositories, not family totals. Qwen alone has more than a dozen repositories above a million a month, so the real Alibaba figure is far higher than any row shown.
None of that changes the shape. Even discounting the phone-sized models and the OCR pipelines, the gap between Chinese and European open-weight adoption is one of order of magnitude, not of degree.
How China Compares to Europe
Our European map found that Europe has one commercial frontier attempt, a handful of maintained national models, two consortia of which one ships, and a growing graveyard. China has eight labs shipping current frontier or near-frontier models, and two dormant tigers. Four of the eight are standalone AI companies and every one of them is listed or preparing to list. Three of the other four sit inside companies that were already public.
The structural difference is not talent and it is not even compute. It is that Chinese labs have a capital route that does not end. A European national model runs until its programme closes. A Chinese lab that reaches the Hong Kong exchange has to answer to shareholders, which is a harder master, but it is a master that keeps paying.
Chinese labs now make Europe’s best argument, open licensing and legal clarity, more convincingly than European ones do. DeepSeek under MIT is more permissive than several models marketed in Europe as sovereign and open.
What to Watch for the Rest of 2026
- Whether DeepSeek actually lists. A STAR market debut at a reported $71bn would make the most influential open-weight lab in the world a public company answerable to Chinese retail investors.
- Whether Moonshot’s licence follows Tencent’s. Kimi K3 is the largest open-weight model anyone has shipped and it carries a bespoke licence. If Moonshot moves to MIT or Apache before its listing, one of the last major custom licenses in Chinese open AI goes with it.
- Whether the tigers come back. 01.AI and Baichuan still have real download bases on old weights. Either could re-enter, and neither has.
- Whether Europe answers on licensing. The one field where Europe can lead without matching Chinese compute is legal clarity, and right now it is losing that field too.
Author: Akos Szima
This article is for information only and is not financial advice. Figures verified against primary sources on 25 August 2026. Download counts are last-30-day figures pulled from the Hugging Face API on 25 August 2026; they are a snapshot and Hugging Face restates them. Valuations described as reported or targeted are flagged as such.
See Also:
European LLMs in 2026: The Complete Map
What Is DeepSeek? China’s Bet Threatening OpenAI
Is Kimi K3 the Model That Breaks Europe’s Sovereign AI Bet?
Frequently Asked Questions
Under OSI-approved licences: DeepSeek V4 and Z.ai’s GLM-5.2 ship under MIT, while Alibaba’s Qwen3.6, ByteDance’s Seed-OSS, Baidu’s ERNIE 4.5 and Tencent’s Hy3 ship under Apache 2.0. Moonshot’s Kimi K3 and MiniMax M2.7 carry custom licences, so open weights in those two cases does not mean open source.
Z.ai, formerly Zhipu, listed in Hong Kong on 8 January 2026 seeking $560m, and MiniMax listed the next day seeking up to $539m and doubled on debut. Alibaba, Tencent and Baidu were already public. Moonshot is preparing a Hong Kong listing on a round valuing it at just over $30bn, and DeepSeek is targeting Shanghai’s STAR market as early as the second quarter of 2027.
Most of them, yes. MIT and Apache 2.0 carry no territorial restriction, which covers DeepSeek, GLM-5.2, Qwen, Seed-OSS, ERNIE 4.5 and Tencent’s Hy3. The exception is Tencent’s older Hunyuan community licence, which states that it does not apply in the European Union, the United Kingdom and South Korea, leaving an EU company running those weights with no grant of rights. Read the licence per release, not per company.
On adoption the gap is one of order of magnitude. Alibaba’s Qwen3-0.6B pulled 25.2 million downloads in the 30 days to 25 August 2026, against 624,507 for Apertus-8B, Europe’s most downloaded open model. China has eight labs shipping current frontier or near-frontier models against Europe’s single commercial frontier attempt.
