Chinese AI Models: 11 Open Picks and 3 License Mistakes
The first version of this page talked about open source Chinese AI models in general terms and even kept template lines such as "[Model Name 1]" where the models should have been. This rewrite replaces all of that with checked facts: I queried the Hugging Face Hub API on 2026-11-08 for the main repositories of DeepSeek, Qwen (Alibaba), Z.ai (GLM), Moonshot AI (Kimi), MiniMax, Tencent and Baidu, read the licenses that are not standard, and ran two small Chinese AI models on a CPU with Ollama 0.40.13. Model lists change every month, so treat the numbers as a snapshot of that date.
Contents
Which Chinese AI models are open on Hugging Face
The Hub API (https://huggingface.co/api/models/<repo>) returns the license tag, the creation date and, for safetensors repos, the total parameter count. This is what it returned for the flagship repos of each lab:
repo license params gated created
deepseek-ai/DeepSeek-V3.2 mit 685.4B False 2025-12-01
deepseek-ai/DeepSeek-V4.1-Flash mit 763.2B False 2026-09-10
deepseek-ai/DeepSeek-R1 mit 684.5B False 2025-01-20
Qwen/Qwen3.8-27B apache-2.0 27.8B False 2026-08-05
Qwen/Qwen3-0.6B apache-2.0 0.8B False 2025-04-27
zai-org/GLM-5.3 other 753.3B False 2026-08-25
zai-org/GLM-5.3-Flash mit 321.3B False 2026-08-25
moonshotai/Kimi-K3 other 2779.9B False 2026-06-13
MiniMaxAI/MiniMax-H3 other 33.1B False 2026-07-28
tencent/Hy3 apache-2.0 298.8B False 2026-07-02
baidu/ERNIE-4.5-21B-A3B-Thinking apache-2.0 21.8B False 2025-09-08
Three things stand out about these Chinese AI models. None of these repos is gated, so anyone can download the weights without an approval step. Most of them use a standard license (MIT or Apache 2.0). And the big ones are big: 685 billion parameters for DeepSeek-V3.2 and almost 2.8 trillion for Kimi-K3 are data-center models, not something for a workstation.

Open source or open weights? Read the license
"Open source" in news about Chinese AI models usually means that the weights are downloadable. What you may do with them depends on the license file, and a license:other tag on the Hub means "read it yourself". I downloaded the four non-standard ones. In short:
- Kimi K3 License (Moonshot AI, 488 words): MIT-style permissions with two conditions. A company that runs a "Model as a Service" business with more than 20 million US dollars of revenue over 12 months "must enter into a separate agreement with Moonshot AI" before commercial use. A product with more than 100 million monthly active users or more than 20 million US dollars monthly revenue must show "Kimi K3" prominently in its interface. Internal use is exempt from both.
- GLM-5.3 License (Z.ai, 395 words): the same structure, but the threshold is 10 billion US dollars of revenue, and the consequence is that the licensee "must pass Z.AI's security review". The smaller GLM-5.3-Flash is plain MIT.
- MiniMax H3 Community License (2671 words): the strictest. Its "Applicable Territory" is "worldwide, excluding the Excluded Territories", defined as "the European Union, the United Kingdom, the Republic of Korea and the United States of America". Products must say "Powered by MiniMax H3", commercial products above 20 million US dollars yearly revenue need written authorization, and military use is listed as prohibited.
- DeepSeek-V3 (2024 release): the code is MIT, the weights use the "DEEPSEEK LICENSE AGREEMENT Version 1.0" with use restrictions in Attachment A (military use among them); the model card says the series "supports commercial use". Later DeepSeek releases in the table are tagged MIT.
The practical rule for all Chinese AI models: check the license file of the exact repo and revision you deploy, not the family name, because the same lab ships MIT and custom licenses side by side (GLM-5.3 vs GLM-5.3-Flash). Keep a copy of the file with your build. If you serve users in the EU, UK, Korea or the US, MiniMax H3 is out unless MiniMax grants a separate license.
Run small Chinese AI models locally with Ollama
You do not need a GPU to try small Chinese AI models from these families. I used the official ollama/ollama image on CPU (setup details in Ollama chatbot on Linux):
podman run -d --name ollama-cn -v ollama-cn-models:/root/.ollama docker.io/ollama/ollama:latest
podman exec ollama-cn ollama pull qwen3:0.6b
podman exec ollama-cn ollama pull deepseek-r1:1.5b
$ ollama list
NAME ID SIZE MODIFIED
deepseek-r1:1.5b e0979632db5a 1.1 GB Less than a second ago
qwen3:0.6b 7df6b6e09427 522 MB 1 second ago
ollama show tells you what you actually got. For Qwen3 0.6B: architecture qwen3, 751.63M parameters, context length 40960, quantization Q4_K_M, capabilities completion, tools and thinking, and the Apache License 2.0. The deepseek-r1:1.5b tag reports architecture qwen2, 1.8B parameters, context length 131072 and the MIT License with "Copyright (c) 2023 DeepSeek". In other words, the small DeepSeek-R1 in Ollama is a distilled model on a Qwen base, not the 684.5B R1 from the table. Read ollama show --license <model> before you ship a quantized copy, because that is the license text Ollama bundles with it.
Chinese and English prompts on a CPU
Qwen3 has a thinking mode that is on by default (ollama show lists "levels false, true" and "default true"). For short answers turn it off with "think": false in the API request. Same question in Chinese and in English, temperature 0:
$ curl -s localhost:11434/api/generate -d '{"model":"qwen3:0.6b","prompt":"用一句话解释什么是 Linux 内核。","think":false,"stream":false,"options":{"temperature":0}}'
response: Linux 内核是操作系统的核心部分,负责管理硬件资源、进程调度和系统运行的底层逻辑。
eval_count: 24 tokens, 52.6 tok/s , load 1.6 s
$ curl -s localhost:11434/api/generate -d '{"model":"qwen3:0.6b","prompt":"Explain in one sentence what the Linux kernel is.","think":false,"stream":false,"options":{"temperature":0}}'
response: The Linux kernel is a core component of the Linux operating system that provides the fundamental software and hardware support for the system, enabling efficient resource management and system functionality.
eval_count: 33 tokens, 51.9 tok/s , load 0.0 s
(The lab sent these requests with a small Python client; the JSON body is the same as in the curl line.) Both answers are correct and fluent, at just over 50 tokens per second on the CPU. In this one sample the Chinese answer also needed fewer tokens (24 against 33), so do not assume that Chinese prompts cost more on these Chinese AI models; measure with your own text.
DeepSeek-R1 distill: reasoning in a separate field
Reasoning models, including the DeepSeek line of Chinese AI models, print a long chain of thought. Ollama returns it in a separate thinking field when you send "think": true, so your application can log it and show only the answer. I asked "Is 221 a prime number? Answer yes or no and give the factors.":
thinking chars: 1160
response (last 300 chars): 221\). - Since the remainder is zero, 221 is **divisible by 13**. ### Step 3: Conclusion Since 221 is divisible by 13 and 17 (other than 1 and itself), it is **not** a prime number. ### Final Answer \[ \boxed{221 \text{ is not a prime number. Its factors are } 1, 13, 17, \text{ and } 221.} \]
eval_count: 899 tokens, 30.2 tok/s
The answer is right (13 × 17 = 221), but it took 899 tokens at about 30 tokens per second, roughly 30 seconds for a one-line question. With both models loaded, ollama ps showed 1.3 GB and 1.0 GB at "100% CPU" and the container used 2.4 GB of RAM. That is the trade-off of reasoning models in practice: better on multi-step problems, slow for chat. The older install story for R1 is in how to install and use DeepSeek R1.
A checklist before you use Chinese AI models in a product
- Pick the exact repo and read its license file. MIT/Apache 2.0 repos are simple; for
license:othercheck revenue, user-count, display and territory clauses as above. - Match the size to your hardware. Total parameters from the Hub are the full download; the 300B to 2.8T models need multi-GPU servers or a hosted API. Small Qwen models and the R1 distills are the Chinese AI models that run on a laptop CPU.
- Test in your own languages and tasks. The two test prompts above say nothing about your domain. Build 20 to 50 real questions and compare.
- Decide where the data goes. Local weights keep prompts on your machine; a hosted API of the same model sends them to the provider, under that provider's terms.
- Pin versions of the Chinese AI models you deploy. Use a fixed Hub revision or Ollama digest (for example
7df6b6e09427) so an update does not change answers silently.
The Hugging Face licenses documentation explains the license tags (for license: other it asks repo owners to add the text in a LICENSE file), and the Ollama API reference covers the think option used here.
FAQ
Are Chinese AI models free for commercial use? Many are: in the snapshot, DeepSeek-V3.2, DeepSeek-R1, the Qwen repos, Tencent Hy3 and Baidu ERNIE 4.5 are MIT or Apache 2.0. Kimi K3, GLM-5.3 and MiniMax H3 add conditions; MiniMax H3 excludes the EU, UK, Korea and the US.
Can I run DeepSeek-R1 on my PC? Not the 684.5B model. The deepseek-r1:1.5b Ollama tag is a small distilled model (qwen2 architecture) and ran on a CPU at about 30 tokens per second.
Do these models send data to China? GGUF weights run by Ollama, as in this lab, are data, not a program. Hub repos that need trust_remote_code run Python code from the repo, so review that code first. Hosted APIs are a different question and depend on the provider's terms.
Where do I find the newest open Chinese AI models? On the organisation pages of the Hub (deepseek-ai, Qwen, zai-org, moonshotai, MiniMaxAI, tencent, baidu), sorted by downloads or date.
Comments
Post a Comment