Skip to main content
Zubnet AILearnWiki › Qwen
Models

Qwen

Also known as: Tongyi Qianwen
Qwen is a family of open-weight large language models developed by Alibaba Cloud. The lineup spans several generations — Qwen 1.5, 2, 2.5, and 3 — with sizes from half a billion to hundreds of billions of parameters, in both dense and mixture-of-experts designs. Most releases ship under the permissive Apache 2.0 license, so they can be run, modified, and deployed commercially without a usage fee.

Why it matters

Qwen is one of the strongest open-weight families available, which makes it a default starting point for teams that want a capable model they can run or fine-tune themselves instead of paying per-token API fees. Its models are especially good at multilingual work, coding, and math, and the spread of sizes means there is usually a variant that fits a given GPU budget. Because the weights are downloadable, a Qwen model can also serve as the base for your own specialized model rather than just an endpoint you call.

Deep Dive

Qwen is not one model but a full product line, and reading the naming scheme makes it much easier to navigate. A typical release name like Qwen2.5-7B-Instruct tells you the generation (2.5), the parameter count (7 billion), and the variant: Instruct means tuned to follow instructions and chat, while Base is the raw pre-trained model meant as a foundation for further training. Each generation ships in several sizes at once, so you can prototype with a small model and scale up without changing tooling. The weights are published on Hugging Face and Alibaba's own ModelScope hub, and they are packaged for every major serving stack, from Ollama on a laptop to vLLM clusters in production.

The Generations and the Sizes

The family has moved through four major generations in quick succession: 1.5, 2, 2.5, and 3. Each generation brought more training data, longer context windows, and better instruction tuning, and each shipped in a ladder of sizes — in the 2.5 line, for example, that meant dense models at roughly 0.5B, 1.5B, 3B, 7B, 14B, 32B, and 72B parameters. The smaller models target laptops and edge devices, the mid-range ones are the workhorses people fine-tune, and the large ones compete with frontier models. Qwen 3 added mixture-of-experts flagships to the mix: the top model has 235 billion total parameters but only activates about 22 billion per token, which keeps inference cost closer to a much smaller dense model.

Qwen 3 also introduced a hybrid thinking mode, where a single model can either answer directly or spend extra compute reasoning step by step before responding, controlled by a flag in the prompt or chat template. That blurred the line between a chat model and a dedicated reasoning model, and it has since become a common pattern across the industry.

The Specialist Variants

Around the core chat models, Alibaba releases specialist lines trained for specific jobs. The Qwen-Coder series targets programming, from autocomplete to agentic coding tasks, and competes directly with dedicated code models. QwQ is the reasoning line, trained to work through math and logic problems with long chains of thought before answering; much of that capability was later folded back into Qwen 3's thinking mode. On the multimodal side, Qwen-VL handles images and documents, and there are audio and math-focused releases as well, plus embedding and reranking models for retrieval pipelines. If a workload has a name, there is usually a Qwen variant aimed at it.

It's Not Just a Chinese Model

A common misconception is that Qwen is mainly useful if you work in Chinese, since it comes from a Chinese company. In practice it is one of the most multilingual families available: the models are trained to handle dozens of languages, the English capability is genuinely strong, and a large share of the people fine-tuning Qwen on Hugging Face never touch Chinese at all. The other worry that comes up is provenance — whether adopting a model from an overseas vendor creates data-exposure risk. With open weights that concern mostly evaporates, because the model runs on your own hardware and no prompt data leaves your machines; the weight file itself is just numbers.

Licensing and Running It Yourself

Most Qwen releases are Apache 2.0, one of the most permissive licenses in the industry, which allows commercial use, modification, and redistribution. A few sizes and research variants have shipped under Alibaba's own licenses with extra conditions, so it is worth checking the model card for the specific release you plan to ship. Day to day, the practical story is simple: quantized GGUF builds run locally through llama.cpp or Ollama, and production deployments typically serve the full-precision or FP8 weights through vLLM or SGLang. Quantization to 4-bit cuts the memory footprint by roughly four times with only a modest quality hit, which is how a 32B model ends up fitting on a single high-end consumer GPU.

Where It Sits in the Open Ecosystem

Qwen competes in the same open-weight arena as Llama, Mistral, Gemma, and DeepSeek, and its releases routinely land near the top of public leaderboards like Chatbot Arena among models whose weights you can actually download. Its deeper influence is as a foundation: the permissive license and strong small models have made Qwen one of the most popular bases for community fine-tuning, and even other labs build on it — DeepSeek released its R1 reasoning model as distillations into Qwen sizes, which says a lot about whose weights the industry trusts. If you are choosing an open-weight base model today, Qwen is on the shortlist almost by default.

← All Terms
ESC