Qwen, short for Tongyi Qianwen, is the family of large language and multimodal models developed by Alibaba Cloud, released both as open-weight models that anyone can download and as hosted proprietary models served through an API and a consumer chat assistant. Among named AI models & assistants, Qwen stands out for the breadth of its open releases, which span tiny models for phones to large mixture-of-experts systems for data centers.
In artificial intelligence practice, Qwen matters because it became one of the most widely fine-tuned open model families, used as a base for thousands of community derivatives, alongside Llama, Mistral, Gemma and DeepSeek.
How the Qwen model family is organized
Qwen is organized as successive generations, each containing several sizes and specialized branches. Every generation typically ships a base model for further training and an instruct model tuned to follow conversational requests.
The following table summarizes the main generations and what each one introduced.
| Generation | Approximate release | Notable change |
|---|---|---|
| Qwen (first generation) | 2023 | First open-weight releases at 7B and 14B parameters |
| Qwen1.5 | Early 2024 | Wider size range and a first mixture-of-experts variant |
| Qwen2 | Mid 2024 | Grouped-query attention across sizes, longer context |
| Qwen2.5 | Late 2024 | Larger pretraining corpus, strong coding and math branches |
| Qwen3 | 2025 | Hybrid thinking modes and flagship mixture-of-experts models |
Alibaba Cloud also offers closed hosted models, such as the Qwen-Max line, that are not released as weights. Release cadence is fast, so the newest generation and its exact sizes should always be checked against the current model cards.
Architecture and training of Qwen models
Qwen models are decoder-only transformers with design choices common to modern open models: rotary position embeddings, SwiGLU activations, RMSNorm and grouped-query attention to reduce memory use during inference. Larger members use a mixture-of-experts design, in which a router activates only a small subset of expert layers for each token.
In a model labeled Qwen3-235B-A22B, for example, the name signals about 235 billion total parameters but only about 22 billion active per token. This gives the knowledge capacity of a very large model at the compute cost of a much smaller one, although all weights must still fit in memory.
The Qwen3 generation introduced a hybrid reasoning mode: the same model can produce an extended chain of thought for difficult problems or answer directly for simple ones, controlled by a switch in the prompt or chat template. Earlier, this behavior required a separate reasoning model such as QwQ.
Pretraining corpora have grown from a few trillion tokens in early generations to tens of trillions of tokens in later ones, with broad multilingual coverage that is particularly strong in Chinese and English.
Specialized Qwen variants
Beyond general chat models, Qwen includes domain and modality branches trained from the same foundations. Choosing the right branch often matters more than choosing the largest size.
| Branch | Modality or domain | Typical use |
|---|---|---|
| Qwen-Coder | Source code | Code completion, repository-level edits, coding agents |
| Qwen-Math | Mathematics | Step-by-step solutions and formal reasoning |
| Qwen-VL | Images, documents and video | Document understanding, chart reading, GUI grounding |
| Qwen-Audio | Speech and sound | Transcription, audio question answering |
| Qwen-Omni | Text, image, audio and video | Real-time spoken assistants with vision |
| Qwen embedding models | Text representations | Retrieval and semantic search |
These branches share tokenizers and much of their pretraining with the general models, which makes fine-tuning across branches predictable. A team that adapts a Qwen chat model can usually reuse the same tooling, chat template and quantization workflow for the coder or vision variant, which lowers the cost of building multi-model systems such as coding agents that read screenshots.
How Qwen compares with other model families
Qwen competes mainly with other open-weight families and is usually evaluated on capability per parameter, licensing and language coverage. Benchmark leadership changes with each release, so comparisons should be read as structural rather than permanent rankings.
| Family | Developer | Typical license | Distinguishing trait |
|---|---|---|---|
| Qwen | Alibaba Cloud | Mostly Apache 2.0 | Very wide size range and many branches |
| Llama | Meta | Custom community license | Large Western ecosystem and tooling |
| DeepSeek | DeepSeek | Permissive, MIT for recent models | Efficient large mixture-of-experts reasoning |
| Mistral | Mistral AI | Mix of Apache 2.0 and commercial | European provider, compact efficient models |
| Gemma | Gemma terms of use | Small models derived from Gemini research |
A practical advantage of Qwen is that most open releases carry the permissive Apache 2.0 license, which allows commercial use and modification with few conditions. Some individual sizes in earlier generations used a custom Qwen license, so each model card must be checked before deployment.
Deploying Qwen and its limitations
Qwen models can be run locally, self-hosted or consumed through Alibaba Cloud’s API. Common open-source runtimes such as vLLM, llama.cpp and Ollama support them, and quantized versions make smaller sizes usable on a laptop.
| Situation | Recommended option | Reason |
|---|---|---|
| Data must stay on premises | Self-hosted open weights | No data leaves the organization |
| Mobile or edge device | Sub-4B model, quantized | Fits limited memory and power |
| Domain-specific assistant | Fine-tuned mid-size instruct model | Good quality per cost after adaptation |
| Highest quality, no hardware | Hosted flagship via API | No infrastructure to manage |
- Content restrictions: models trained in China follow domestic content regulations, which can cause refusals or filtered answers on politically sensitive topics.
- Benchmark versus real performance: strong leaderboard scores do not guarantee equal quality in every language or domain.
- Memory requirements: mixture-of-experts models are cheap per token but still need memory for all weights.
- Governance review: some organizations restrict models by country of origin, so procurement policy should be checked early.
In simple terms
Qwen is a set of AI models made by Alibaba, the Chinese technology company. It can chat, write, translate, program, and in some versions look at pictures or listen to speech.
What makes Qwen special is that most versions are free to download and run on your own computer. It comes in many sizes, from small ones for a phone to very large ones for a server farm, much like a car maker offering everything from a city car to a truck. You simply pick the size that suits the job.
A practical example
A mid-sized logistics company wants an internal assistant that answers questions about shipping contracts written in English, German and Chinese. Its legal team does not allow contracts to be sent to external AI services.
The engineering team downloads a mid-size Qwen instruct model under the Apache 2.0 license, runs it with vLLM on two on-premises GPUs, and fine-tunes it on about 3,000 internal question-and-answer pairs. A Qwen embedding model indexes the contract archive for retrieval.
After a four-week pilot, staff find relevant clauses in seconds instead of searching manually, and no document ever leaves the company network. The team keeps the option to switch to a newer Qwen generation later without changing its serving stack.