Qwen (全能问答, short for Tongyi Qianwen — “to comprehend a thousand questions”) is the AI model family developed by Alibaba Group’s cloud division. Since its public launch in September 2023, it has grown into the world’s most widely deployed open-weight AI series, covering language, code, mathematics, vision, audio, and video generation.
What Qwen is
Qwen is not a single model but a family spanning dozens of models at different sizes and specializations. The core offering is a series of large language models (LLMs) — the same kind of AI that powers ChatGPT and Claude — but most Qwen models are released under open licenses that let anyone download, modify, and deploy them without paying Alibaba.
The name comes from the Chinese 通义千问 (Tongyi Qianwen), meaning “to comprehend a thousand kinds of questions.” Alibaba’s consumer chatbot — which runs on Qwen models — is called Qianwen in China; internationally, the model family is branded as Qwen.
The model family
Qwen has grown into one of the most complete AI model families available:
| Specialty | Key models |
|---|---|
| General language | Qwen3 (0.6B – 235B params), Qwen2.5-72B |
| Coding | Qwen2.5-Coder, Qwen3-Coder |
| Mathematics | Qwen2.5-Math, QwQ-32B |
| Vision and images | Qwen2.5-VL, Qwen3-VL |
| Audio / speech | Qwen3-ASR |
| Multimodal (text + image + audio + video) | Qwen2.5-Omni |
The latest open-weight generation, Qwen3 (released April 2025), was trained on 36 trillion tokens across 119 languages. Its largest open model — Qwen3-235B — uses a mixture-of-experts architecture, activating only 22 billion parameters at a time, which makes it far cheaper to run than a dense 235B model would be. Smaller Qwen3 models (4B, 8B) match or exceed models five times their size from earlier generations.
As of mid-2026, the latest proprietary variant, Qwen3.6-Plus, rivals Claude Opus 4.6 and GPT-5 on agentic coding benchmarks, scoring 78.8% on SWE-bench Verified.
Licensing — is it open source?
Most Qwen models are genuinely open source under the Apache 2.0 license — the most permissive mainstream AI license. Apache 2.0 permits unlimited commercial use with no usage reporting or royalties, making it more permissive than Meta’s Llama licenses. All Qwen3 models (0.6B through 235B) and most Qwen2.5 sizes are Apache 2.0.
A small number of larger models carry stricter terms (Qwen2.5-72B uses a custom Qwen License), and the most capable variants (Qwen3.6-Plus, Qwen3.7-Max) are API-only. But the open tiers are extensive: over 100 open-weight Qwen models are on Hugging Face, with more than one billion cumulative downloads — overtaking Meta’s Llama as the most-downloaded model family globally in late 2025.
How to use Qwen
Try it free in a browser: qwen.ai (Qwen Studio) offers a free chat interface with no sign-up required; creating a free account unlocks a higher daily message limit.
Run it locally: Qwen models are available at huggingface.co/Qwen and can be run on a personal computer via Ollama, a tool that manages local AI models. Qwen3-8B runs well on a modern laptop with 16 GB of RAM.
Use the API: Alibaba Cloud’s Model Studio provides OpenAI-compatible API endpoints for Qwen models, including proprietary tiers. Pricing is token-based, similar to other AI providers; as of June 2026, check the Alibaba Cloud Model Studio for current rates, as the free tiers and per-token costs update frequently.
In the news
Today’s news covers two stories directly tied to Qwen: Alibaba released Qwen-AgentWorld, a simulation environment for testing AI agents before real-world deployment, and Anthropic accused Alibaba of running a large-scale distillation attack on Claude to transfer its capabilities into Qwen models. Both reflect how central Qwen has become to the global AI race.
FAQ
Is Qwen better than ChatGPT?
For coding and structured reasoning tasks, Qwen3’s top models match or exceed GPT-4o. ChatGPT still leads for general conversational polish in some languages, but the gap has narrowed substantially.
Can I use Qwen commercially?
All Qwen3 models and most Qwen2.5 sizes are under Apache 2.0, which permits unrestricted commercial use. Qwen2.5-72B and the proprietary API models carry separate terms — check the license file on each Hugging Face model card.
Does Qwen support Georgian?
Yes. Qwen3 was trained on 119 languages including Georgian. The web demo at qwen.ai accepts Georgian text input and responds in Georgian.
What is the difference between Qwen and Tongyi Qianwen?
They are the same thing: Tongyi Qianwen (通义千问) is the full Chinese name; Qwen is the international brand name used outside China.
Sources: Wikipedia — Qwen · Qwen3 technical blog · Hugging Face — Qwen