Glossary
Qwen: a family of language models
In short
Qwen is a family of language models from the Qwen team at Alibaba Cloud. Many versions are published with open weights under the Apache 2.0 license.
Models and products · Qwen
MMOOZZGG.com editorsChecked: 1 min read
For a child
Imagine a building set that the maker lets everyone take home for free. You can build a toy car with it in your room, without giving the pieces to anyone, and even change it your own way. Some smart helper programs are shared the same way: you download one and use it on your own computer. But such a helper can make mistakes too, so adults check what it made up.
In plain words
These are large language models made by the Qwen team at Alibaba Cloud. They write and translate texts, answer questions, help with code and understand more than a hundred languages. The main feature: you can download the files of many models and run them yourself, without sending data to someone else’s cloud. The line has small versions for a laptop and large ones for a server with several graphics cards. For example, a shop puts the model on its own server, and it answers buyers about delivery without sending the chats abroad.
For business
For a company in Russia the license is the main point. The technical report and the model cards name Apache 2.0, which allows commercial use, modification and running on your own hardware. Customer chats stay inside your network, so the rules of 152-FZ, Russia’s personal data law, are easier to meet. Costs come from servers with graphics cards, setup and support, not from a fee for every request. The Alibaba Cloud API is a separate service with its own contract and terms, and a lawyer checks them before you connect. Before rollout, test the quality of answers in Russian on your documents and the load on the hardware: a pilot measures both.
For a specialist
The Qwen3 line (technical report by Yang and co-authors, May 2025) includes dense models and Mixture of Experts (MoE) models from 0.6 to 235 billion parameters. According to its model card, the flagship Qwen3–235B-A22B has 235 billion parameters, and 22 billion of them are active for each token. The native context is 32,768 tokens, extended to 131,072 with YaRN scaling. A thinking mode and a fast mode without reasoning are combined in one model. You switch through the chat template, and a reasoning budget limits the computing per answer. The number of languages and dialects grew from 29 in Qwen2.5 to 119. The Qwen3–2507 update came as separate Instruct and Thinking versions. For server inference, the model card names vLLM and SGLang; locally, Ollama, LM Studio and llama.cpp support the models. The weights are published under Apache 2.0.
Example
A manufacturer runs an open Qwen model on its own server and connects it to an archive of process cards. Technologists ask about tolerances and machining modes, and the drawing files never leave the factory network.
Related terms
Questions and answers
Sources
Not sure whether your company needs this?
AI solutions map