A year ago most people outside the industry could name one Chinese AI company: DeepSeek. Today there are at least nine Chinese developers shipping models that sit close to the global frontier, and several of them give their models away for anyone to download. This guide explains who they are, what they have released as of late November 2026, why chips and US export rules matter so much, and where the UK fits in.
Why Chinese models matter
The big American labs mostly keep their best models closed: you can use them through an app or an API (a paid connection for developers), but you cannot download them. Many Chinese labs take the opposite approach and publish "open weights", the trained files that let anyone run and modify a model on their own computers. That makes them cheap to use and easy to build on, which is why they now turn up inside products all over the world.
Two caveats apply throughout. First, almost every benchmark score below comes from the company that made the model. Second, "open" rarely means a standard open-source licence: most labs attach their own terms, especially for very large commercial users.
The independent labs
DeepSeek
The Hangzhou lab that shocked markets in early 2025 is still the reference point. Its newest model, DeepSeek-V4.1-Flash-1017, arrived on 17 October as what DeepSeek called a "minor update" that finishes tasks faster with shorter answers. The weights are on Hugging Face under the permissive MIT licence, and the model handles up to a million tokens of context (roughly the text of several long books at once). DeepSeek has also just completed a second funding round of about 100 billion yuan (about $15 billion) and is preparing a Shanghai stock market listing.
Moonshot AI (Kimi)
Beijing-based Moonshot makes the Kimi assistant. Its flagship, Kimi K3.1 Pro, is a 2.8-trillion-parameter model (parameters are the internal settings a model learns) that activates about 104 billion of them for each word it processes. Moonshot published the weights on 4 November under its own Kimi K3.1 Pro licence rather than a standard open-source one.
Z.ai (GLM)
Z.ai, formerly Zhipu AI, released GLM-5.3 in August with an unusual warning: its ability to find and exploit software flaws had grown faster than expected, so the company held back the weights for two weeks of safety work before publishing them. Z.ai is listed in Hong Kong.
MiniMax
MiniMax, also listed in Hong Kong, focuses on cheap, fast models. MiniMax M3.1 Flash, launched on 19 October, uses about 23 billion of its roughly 428 billion parameters per token and costs $0.15 per million input tokens and $0.60 per million output tokens on MiniMax's API. Investors have been less enthusiastic than developers: on 18 November both MiniMax and Z.ai shares were about 90% below their highs of the past year.
StepFun
Shanghai's StepFun launched Step 5 Preview in September, a 600-billion-parameter model with 27 billion active per token, pitched at coding and financial analysis. It promised open weights for 15 October and has since published them on Hugging Face.
The tech giants
Alibaba (Qwen)
Alibaba has released many of its Qwen models with open weights, but its newest flagship breaks the pattern. Qwen4-Max, launched on 16 November, has 2.4 trillion parameters and a million-token context, and its weights are closed. Alibaba says smaller Qwen4 models will be released openly.
ByteDance (Doubao and Seed)
TikTok's owner builds its Doubao models through its Seed research team and puts them into consumer and office apps. Its Doubao Work agent ran only ByteDance's own Doubao 2.1 models until 12 November, when it added Kimi K3.1 Pro, GLM 5.3 and DeepSeek V4.1 Flash, a sign that even the biggest platforms now mix in rivals' models.
Tencent (Hy, formerly Hunyuan)
Tencent launched its Hy4 model on 4 November and published the weights on 20 November. It is a 770-billion-parameter model with 49 billion active per token. Tencent's own model card admits "a notable gap remains" between Hy4 and the best closed models.
Baidu (ERNIE)
Baidu used its annual Baidu World event on 12 November to promote ERNIE 5.5 Flash Preview, a 280-billion-parameter model it says will have open weights by 25 November. Baidu also designs its own AI chips through its Kunlunxin business, which has started shipping an M100 inference chip.
Chips and US export controls
Training a frontier model needs tens of thousands of advanced chips, and the best ones are designed by American firms, chiefly Nvidia. Since 2022 the US has restricted sales of its most powerful AI chips to China. Chinese labs worked around this partly by renting computing power in data centres in other countries.
That route has now narrowed sharply. On 16 November the US Commerce Department's Bureau of Industry and Security issued a rule, effective 19 November, requiring licences for advanced chips supplied to data centres anywhere in the world that provide large amounts of computing to companies headquartered in China. It also added licence requirements for Indonesia, Malaysia, Singapore, Thailand and Brazil. A day later Nvidia said it was no longer shipping its licensed Hopper-generation chips, including the H200, to China.
Chinese companies say they will lean harder on home-grown chips, such as those Baidu designs through Kunlunxin. Tencent's president, Martin Lau, told analysts on 18 November that the new US rules would slow the next round of Chinese models in the short run and push the whole industry towards domestic supply chains.
The UK angle
The UK's AI Security Institute (AISI), the government's specialist testing body, has published its first public comparison of open and closed models on cyber security. In July it found that GLM-5.2 and DeepSeek V4-Pro performed roughly like closed frontier models released four to seven months earlier, a narrower gap than in 2025. It also noted that the safeguards on these open models were easy to get round. In a parliamentary answer on 9 November, the government said AISI "does not give a running commentary" on the models it tests, but that open-weight testing remains a priority.
Licensing has a UK dimension too. Some of Tencent's earlier open models came with a licence that expressly did not apply in the UK or the EU. The new Hy4 licence grants rights worldwide, but large companies selling access to the model as a service need Tencent's written permission.
How to read the claims
Treat launch-day benchmark tables as marketing until independent testers report. Check the licence before building a product on any open model, and check where your data would be processed if you use a Chinese company's own app or API. For many British businesses the practical route is to run open weights on infrastructure they already trust, or to use them through a major cloud provider.
FAQs
Which Chinese AI models can I download and run myself?
As of late November 2026, DeepSeek-V4.1-Flash-1017, Kimi K3.1 Pro, GLM-5.3, MiniMax M3.1 Flash, Step 5 Preview and Tencent's Hy4 all have weights on Hugging Face, each under its own licence. Alibaba's flagship Qwen4-Max does not, and Baidu says ERNIE 5.5 Flash Preview weights are due by 25 November.
What did the US change on chips in November 2026?
A Commerce Department rule issued on 16 November and effective on 19 November requires licences for advanced US chips supplied to data centres anywhere that provide large amounts of computing to China-headquartered companies, and adds licence requirements for Indonesia, Malaysia, Singapore, Thailand and Brazil. Nvidia also said it had stopped shipping licensed Hopper chips, including the H200, to China.
Has the UK government tested Chinese AI models?
Yes, in part. The AI Security Institute published cyber test results in July 2026 for GLM-5.2 and DeepSeek V4-Pro, finding they trailed closed frontier models by four to seven months. Ministers say open-weight testing remains a priority but that AISI does not comment on every model it tests.
Sources
- Hugging Face: DeepSeek-V4.1-Flash-1017 model card
- Caixin Global: DeepSeek Closes 100 Billion Yuan Funding Round (20 Nov 2026)
- Hugging Face: Kimi K3.1 Pro model card (Moonshot AI)
- Kimi.ai on X: Releasing the model weights of Kimi K3.1 Pro (4 Nov 2026)
- Z.ai: GLM-5.3: Frontier Coding with Emergent Cyber Capabilities (14 Aug 2026)
- Z.ai on X: GLM-5.3 is now open-weight (28 Aug 2026)
- Hugging Face: MiniMax-M3.1-Flash model card
- MiniMax: MiniMax M3.1 Flash launch post (19 Oct 2026)
- Investing.com: ZAI vs. MiniMax share performance (17 Nov 2026)
- StepFun: Step 5 Preview: Advancing the Pareto Frontier
- Hugging Face: Step 5 Preview weights (StepFun)
- Alibaba Cloud: Alibaba Unveils Qwen4-Max (16 Nov 2026)
- EastFrontier: ByteDance's Doubao Work Opens to DeepSeek, Kimi and GLM Models (13 Nov 2026)
- Hugging Face: Tencent Hy4 model card
- Baidu via PR Newswire: Baidu Unveils Latest ERNIE Models (12 Nov 2026)
- Federal Register: Addressing Remote Access to Advanced Semiconductors (19 Nov 2026)
- Reuters via Investing.com: Nvidia stops shipping US-approved AI chips to China (18 Nov 2026)
- Lianhe Zaobao: Tencent president says new US AI chip controls will slow China's model development (19 Nov 2026)
- AI Security Institute: How Far Behind the Frontier are Leading Open Weight Models on Cyber? (17 Jul 2026)
- UK Parliament: Written answer on the AI Security Institute, UIN 37340 (9 Nov 2026)
- Hugging Face: Tencent Hunyuan-A13B licence (territory excludes EU, UK and South Korea)
Follow AI Decoded on Substack
Weekly plain-English AI news and verified UK events.
Subscribe