Open source AI models: China, the USA, and who else is worth knowing.
Some of the world's most capable AI models are free to download and run on your own hardware, and most of the biggest now come from China. Here is who makes them, the sizes they come in, and what to check before your business uses one.
The short answer
An open-weight AI model is one whose trained model you can download, run and adapt yourself, usually for free.
Chinese labs such as Alibaba (Qwen), DeepSeek, Moonshot (Kimi) and Z.ai (GLM) now release the largest and most downloaded open models, while US labs such as Google, OpenAI, Meta and NVIDIA lead on smaller models that run on a laptop.
Before you use one, check its licence, pick a size your hardware can run, and treat it like any other AI tool in your policy.
Key facts
- Chinese models made up 41 percent of downloads on Hugging Face, the main library for open models, in the past year, overtaking the US in both monthly and overall downloads for the first time.Hugging Face, 17 March 2026 · source
- In almost every month of 2026, the largest open model from a Chinese lab was bigger than any model an American lab released. US releases stayed under 130 billion parameters in five of the first seven months.Hugging Face, 14 August 2026 · source
- 151,448 models on Hugging Face are built on Alibaba's Qwen, 2.6 times as many as are built on Meta's models.Hugging Face, 14 August 2026 · source
- Models under 1 billion parameters account for 83 percent of all-time downloads, and models over 100 billion for 1 percent.Hugging Face, 14 August 2026 · source
- In a US government evaluation, agents built on DeepSeek R1-0528 were 12 times more likely than US frontier models to follow malicious instructions.NIST CAISI, 30 September 2025 · source
What does open source mean for an AI model?
Most "open source" AI models are really open-weight: you get the finished model to download and run, but not the data or all the code used to train it.
A large language model is, at heart, a very large file of numbers called weights, learned during training. An open-weight model publishes that file. You can run it on your own computer or server with no internet connection, fine-tune it on your own material, and build it into your own products. A closed model, such as ChatGPT, Claude or Copilot, only runs on its maker's servers, and you reach it through an app or an API.
Fully open source AI goes further and also publishes the training data and code, so anyone can see how the model was made. That is rarer. In everyday use, people say "open source" for both, and this guide does too.
Open does not always mean free to use however you like. The licence sets the rules, and they vary a lot from model to model.
Who makes the leading open models?
Four Chinese labs lead at the top end, four US companies lead on smaller models, and a handful of others are worth knowing.
| Lab | Notable open models | Sizes | Licence |
|---|---|---|---|
| China | |||
| Alibaba (Qwen) | Qwen3.5, Qwen3.6, Qwen3.8 | 0.8B to 2.4T | Apache 2.0 for most; check the largest |
| DeepSeek | DeepSeek V4 Pro, V4.1 Flash | About 760B to 1.7T | MIT |
| Moonshot AI | Kimi K2.7-Code, Kimi K3 | 1T to 2.8T | Modified MIT; Kimi K3 licence |
| Z.ai | GLM-5.3 | 753B | Custom |
| USA | |||
| Gemma 4 | 2B to 31B | Apache 2.0 | |
| OpenAI | gpt-oss-20b, gpt-oss-120b | 21B and 117B | Apache 2.0 |
| Meta | Llama 4 Scout, Maverick | 109B and 400B | Llama 4 Community Licence |
| NVIDIA | Nemotron 3.5 Super, Nemotron 3 Ultra | About 120B to 560B | OpenMDW and NVIDIA licences |
| Elsewhere | |||
| Mistral AI (France) | Mistral Small 4, Mistral Large 3 | 119B and 675B | Apache 2.0 |
| Microsoft | Phi-4 family | Around 15B | MIT |
| IBM | Granite 4.0 | Up to about 32B | Apache 2.0 |
Sizes are total parameters (B = billion, T = trillion). Licences and sizes from each lab's model pages on Hugging Face, checked 8 November 2026. Other Chinese labs, including MiniMax, Tencent and Xiaomi, also release large open models.
Two patterns stand out. China's top labs release their best models openly, at sizes that need a server to run. The US labs release smaller, efficient models, while keeping their most capable ones, such as GPT, Claude and Gemini, closed. Our AI tools comparison covers the closed assistants.
Why is China ahead on open models?
Chinese labs give away their frontier models; most US labs give away smaller ones and sell access to the best.
Hugging Face's analysis of 2026 releases found China's largest open model each month ranged from 754 billion to 2.78 trillion parameters, while US labs mostly stayed under 130 billion. A few US releases, such as NVIDIA's Nemotron 3 Ultra and Arcee's Trinity-Large, are the exceptions.
Scale on the download charts follows from that. Qwen comes in every size from under 1 billion parameters upward, so developers build on it: more models on Hugging Face now start from Qwen than from any other family. Hugging Face counts downloads by the lab that made the model, not where the person downloading it lives, so the figures measure who builds the models people use, not who uses AI most.
The terms are tightening at the top, though. Some of the newest Chinese giants, including Kimi K3, come with more licence conditions than earlier releases, so check each one rather than assuming it is free to use.
Is it safe to use a Chinese AI model?
Running the weights on your own machine does not send your data to China, but the model's behaviour comes with it, so test it before you rely on it.
Separate two different risks. The first is where your data goes. Using DeepSeek's or Kimi's own app or API sends your prompts to servers under Chinese law, which is a data question like any overseas service (see our guide to AI and the Privacy Act). Downloading the weights and running them yourself, on your own hardware or a New Zealand server, sends nothing anywhere.
The second is how the model behaves, wherever it runs. In September 2025 the US government's Center for AI Standards and Innovation tested three DeepSeek models against US models. It found the DeepSeek models echoed four times as many inaccurate and misleading Chinese Communist Party narratives, and that agents built on them were far easier to hijack with malicious instructions. That is one government's evaluation of one family of models at one point in time, but it is a fair warning: censorship on some topics and weaker safety guardrails travel with the weights.
The practical rule is the same for any model, from any country. Use local open models for internal, lower-risk work, test them on your own tasks, and keep a person checking anything that reaches customers or takes actions. See our guide to AI agents for why that matters more once a model can act.
What sizes do open models come in?
Size is measured in parameters, and it decides how much memory a model needs and so what hardware can run it.
Parameters are the model's learned settings. More generally means more capable, but also more memory. Many large models now use a mixture-of-experts design, which uses only part of the model for each word. OpenAI's gpt-oss-120b has 117 billion parameters but uses 5.1 billion at a time, so it runs fast. It still needs memory for all 117 billion.
| Size | Examples | Runs on |
|---|---|---|
| Tiny, under 4B | Gemma 4 E2B, Qwen3.5 2B | Phones and any recent laptop |
| Small, 8B to 35B | gpt-oss-20b, Gemma 4 31B, Qwen3.6-27B | A good laptop or a Mac mini |
| Medium, 70B to 130B | gpt-oss-120b, Nemotron 3.5 Super, Mistral Small 4 | A 64GB to 128GB workstation |
| Frontier, 500B and up | DeepSeek V4, GLM-5.3, Kimi K3 | Servers with several data-centre GPUs, or a hosted API |
Most businesses that try open models start with the small band: they are good enough for drafting, summarising and answering questions over documents, and they run on hardware you may already own. Our guide to running AI on your own computer covers the hardware and setup step by step.
What should you check in the licence?
Three things: whether commercial use is allowed, whether there are size or revenue thresholds, and whether you must credit the model.
- Permissive licences. Apache 2.0 and MIT, used by Gemma 4, gpt-oss, most of Qwen, DeepSeek and Mistral, allow commercial use with few conditions beyond keeping the licence notice.
- Thresholds. Meta's Llama 4 licence requires a separate licence for companies whose products had more than 700 million monthly users. The Kimi K3 licence requires a separate agreement for businesses that sell model access as a service and earn more than US$20 million a year, and it exempts internal use.
- Attribution. Some licences require you to name the model in your product above a certain size, as Kimi K3's does above 100 million monthly users.
For most New Zealand businesses using a model internally, these thresholds will never apply. But read the licence file on the model's own page before you build anything you sell, and record the model and licence in your AI register.
Our view.
Open models are a tool for specific jobs, not a replacement for a frontier assistant. A small model on your own hardware is ideal for sensitive drafting, summarising internal documents and working offline. For complex reasoning and agent work, the closed frontier models are still ahead.
Country of origin matters less than people think once the weights run on your own machine. What matters is how the model behaves on your work, which you only find out by testing it.
And a model downloaded onto a laptop without anyone knowing is shadow AI, however free it was. Put open models through the same approval and register as every other tool. Governance is what lets you use them with confidence.
What are the risks of open models?
Five come up most often, and each has a simple guardrail.
- Untested behaviour. Censorship, weak safety guardrails or plain errors. Test every model on your own tasks before you rely on it, and keep a person reviewing outputs. See why AI models hallucinate.
- Tampered downloads. Anyone can upload a copy of a model. Download only from the lab's official page.
- Licence surprises. Read the licence file, not a summary, before you build a product on a model.
- No vendor support. Nobody patches or updates a model you run yourself unless you do. Assign an owner.
- Shadow AI. Free and offline also means invisible to IT. Add open models to your approved tools list.
Where should you start?
Know where you stand, then pick one job an open model could do better than your current tools.
- Find out where you stand. The free NZ AI Benchmark is 40 questions, takes about 20 minutes, and shows where your organisation stands and the next move.
- Name the job and the reason. Privacy, cost or working offline. If none applies, a closed assistant is probably simpler.
- Pick a model from the lab's official page and check its licence.
- Test it on real, non-sensitive work against the tool you use today. Our guide to running AI locally has the setup steps.
- Add it to your AI policy and register. Our AI policy guide covers who may approve new tools.
Frequently asked questions.
Are open source AI models free?
The models themselves are usually free to download. You pay for the hardware to run them, or for a cloud provider to host them, and for the time to set them up and keep them updated. Licences for some models add conditions for very large companies.
Can New Zealand businesses use Chinese AI models?
Yes. There is no New Zealand law against it. Running the weights on your own hardware keeps your data in-house. Using a Chinese provider's own app or API sends data offshore, so treat that like any overseas service under the Privacy Act. Either way, test the model's behaviour before you rely on it.
Does a local AI model send my data anywhere?
Not if it runs entirely on your own hardware. The model works offline once downloaded. Some apps that run models also offer cloud options, so check the settings and switch them off if you want everything to stay on your machine.
What is the difference between open source and open-weight?
Open-weight means the trained model is published for anyone to download and run. Open source, strictly, also publishes the training data and code. Most models called open source are open-weight.
Which open AI model is best?
It depends on your hardware and your task. For a laptop or Mac mini, Google's Gemma 4, OpenAI's gpt-oss-20b and Alibaba's smaller Qwen models are strong starting points. At the top end, DeepSeek, Kimi, GLM and Qwen lead, but they need a server. Test two or three on your own work.
Are open models as good as ChatGPT or Claude?
The largest open models come close on many tasks. The small ones you can run on a laptop are good for everyday drafting and summarising, but behind the leading closed models on complex reasoning, long documents and agent work.
Sources.
- State of Open Source on Hugging Face: Spring 2026, Hugging Face (17 March 2026; accessed 8 November 2026)
- State of Open Models: Summer 2026 Observations, Hugging Face (14 August 2026; accessed 8 November 2026)
- CAISI evaluation of DeepSeek AI models finds shortcomings and risks, NIST (30 September 2025; accessed 8 November 2026)
- Qwen models, Alibaba (Hugging Face; checked 8 November 2026)
- DeepSeek models, DeepSeek (Hugging Face; checked 8 November 2026)
- Kimi K3 licence, Moonshot AI (checked 8 November 2026)
- GLM models, Z.ai (Hugging Face; checked 8 November 2026)
- Gemma 4 model overview, Google (checked 8 November 2026)
- gpt-oss-120b model card, OpenAI (checked 8 November 2026)
- Llama 4 Community License Agreement, Meta (checked 8 November 2026)
- Nemotron 3.5 Super model card, NVIDIA (checked 8 November 2026)
- Mistral Small 4 model card, Mistral AI (checked 8 November 2026)
Related topics
Keep going.
Your next step
Work out where open models fit in your business.
The free NZ AI Benchmark takes about 20 minutes and shows where your organisation stands. Or book a 45 minute Discovery Call with Justin or Ben to talk through the right mix of open and closed AI tools for your work.
Changelog: 8 November 2026, first published.