Chinese open-source AI models: what they are and how to try them (Qwen, DeepSeek, Kimi, GLM, MiniMax)

Qwen, DeepSeek, Kimi, GLM and MiniMax are five families of AI models built by Chinese companies that publish their "weights", the files that let anyone download and run the models. As of November 2026 most of them are huge and used online, but one (Qwen) also has a version that fits on a powerful PC.

This guide explains what they are, which licenses they use, how to try them and, above all, what changes for your data between online and local use. Performance figures are those declared by the makers on their official model cards and have not been verified by third parties.

What does "open source" mean for these models?

For an AI model it usually means "open weights": you can download the model, use it and often adapt it. That does not mean everything is equally free, because each model card names a license. Some are standard (MIT, Apache 2.0), others are custom licenses with conditions, for example for anyone offering the model as a paid service. Before using a model at work, read the license on its official page.

What are the latest versions as of November 2026?

These are the most recent models read on each maker's official Hugging Face page, with the data declared on the model cards:

FamilyRecent modelDeclared sizeLicense (as listed)On Ollama
QwenQwen3.8-27B27 billion parametersApache 2.0Yes, local: qwen3.8 (18 GB)
DeepSeekDeepSeek-V4.1-Flash-1017763 billion totalMITYes, but cloud only
Kimi (Moonshot AI)Kimi-K3.1-Pro2.8 trillion total, 104 billion activeCustom "Kimi K3.1 Pro License"Yes, but cloud only
GLM (Z.ai)GLM-5.3 (753 billion) and GLM-5.3-Flash (320 billion, 18 active)753 and 320 billionGLM-5.3: custom permissive-style license; GLM-5.3-Flash: MITYes, but cloud only
MiniMaxMiniMax-M3.1-Flashabout 428 billion, 23 active"minimax-community" licenseYes, but cloud only

Qwen also has larger models: Qwen3.8-2.4T-A95B (2.4 trillion total, 95 billion active, "qwen3.8-max" license) and Qwen3.8-Flash-Next (125 billion total, 6 active, "qwen-community-1.0" license). "Mixture-of-experts" models have many parameters in total but activate only a part of them for each answer. The DeepSeek, Kimi and MiniMax cards list a context window of up to about 1 million tokens.

What licenses do they have?

  • Apache 2.0 and MIT: standard, permissive licenses. Qwen3.8-27B is Apache 2.0; DeepSeek-V4.1-Flash-1017 and GLM-5.3-Flash are MIT.
  • Kimi K3.1 Pro: free use, with two conditions stated in the text: an agreement with Moonshot AI if you offer the model as an API service above US$20 million in revenue over 12 months, and a requirement to display the model name for products above 100 million monthly users or US$20 million in monthly revenue.
  • GLM-5.3: a permissive license with a clause for very large service providers (over US$10 billion in revenue).
  • MiniMax M3.1-Flash: free for non-commercial purposes; commercial use requires the notice "Built with MiniMax M3.1-Flash", a notification to MiniMax and, above US$20 million in annual revenue, written authorization. The text also lists prohibited uses.

These are summaries: for professional use, read the full LICENSE file on the model card you choose.

Which ones can run on your own PC?

Of the pages read, only Qwen has a PC-sized local version on Ollama: qwen3.8:27b, an 18 GB download with a 256K context window that accepts text and images. The Ollama library also has smaller Qwen models (Qwen3.5 from 0.8 to 9 billion parameters) and the smaller versions of DeepSeek-R1, an older model, from 1.5 to 70 billion. Exact memory requirements for your graphics card are not given on the pages read: check the official Ollama site.

The recent DeepSeek, Kimi, GLM and MiniMax models on Ollama carry the "cloud" label: they run on Ollama's servers, not on your computer. Actually downloading and running their weights takes very powerful machines, like those of companies and data centers.

What changes for privacy: app, API or local?

  • The maker's app or website, and its official API: the text you write is sent to the provider's servers. How your data is handled depends on its privacy policy, which is worth reading before you enter work documents or client data. We could not read each maker's policy, so we make no claims about them.
  • "Cloud" models on Ollama: they are used with an Ollama account and run on Ollama's infrastructure. Ollama states in its blog that its cloud "does not retain your data".
  • Local weights: if you download a model and run it on your own computer, the data you give it stays on that computer, because no online service receives it. That is the case for qwen3.8:27b on Ollama.

How do you try them?

With Ollama installed, these are the commands shown on the library pages. The first downloads 18 GB and runs on your PC; the second runs in Ollama's cloud.

ollama run qwen3.8:27b
ollama run glm-5.3-flash:cloud

Alternatively you can use the makers' apps or APIs, checking prices and privacy policy on their sites first.

What does this mean for people who work?

  • For confidential data (clients, contracts, drafts), a local model such as Qwen3.8-27B is better than an online service.
  • For heavy tasks, the big models are in practice used through an API or the cloud: weigh cost and privacy policy.
  • If you sell a service built on these models, read the license: conditions differ from model to model.

Frequently asked questions

Are Chinese open-source models free?

The weights can be downloaded for free, under each license's conditions. Using the maker's app, API or cloud may cost money: check prices on the official sites.

If I run the model locally, does my data leave my PC?

If the model runs on your computer, the data stays on the computer. Be careful with Ollama models labeled "cloud": they run online.

Which one should I try first?

For a powerful PC, Qwen3.8-27B on Ollama is the only one of those read that combines an Apache 2.0 license with direct local execution. The others are tried online.

Official sources

This article was prepared with the help of artificial intelligence tools based on the official documentation, and is reviewed before publication.

Condividi questo articolo sui tuoi socialShare this article on your social networksComparte este artículo en tus redes socialesTeile diesen Artikel in deinen sozialen Netzwerken

No products were found matching your selection.

Iscriviti alla newsletter di You vs AISubscribe to the You vs AI newsletterSuscríbete al boletín de You vs AIAbonniere den Newsletter von You vs AI