Qwen (Tongyi Qianwen) Development History: From First Release to Qwen3
Qwen (Tongyi Qianwen) is Alibaba's family of large language models, developed by Alibaba Cloud and the DAMO Academy. Since its public launch in 2023, it has gone through multiple generations and become one of the most influential open-source and commercial LLMs in the region. This article outlines key milestones, open-source releases, and product evolution.
2023: The Birth of Tongyi Qianwen
In April 2023, Alibaba Cloud officially launched the Tongyi Qianwen model under the English brand name Qwen and opened a limited web trial. In August, Qwen-7B was released as open source. In September, Tongyi Qianwen opened to the public and the Tongyi app launched, led by then Alibaba Cloud CTO Zhou Jingren.
| Date | Event |
|---|---|
| April 2023 | Tongyi Qianwen launch and web trial |
| August 2023 | Qwen-7B open source |
| September 2023 | Public availability and Tongyi app launch |
2024: Qwen2 and Broad Open-Source Release
Qwen2 Series (June 2024)
Qwen2 was released as open source with multiple scales: 0.5B, 1.5B, 7B, 72B, and 57B-A14B (MoE). It supports up to 128K tokens of context and uses optimizations such as GQA, with significant gains on multilingual, code, and math benchmarks.
Qwen2.5 (September–October 2024)
Qwen2.5 was described as the largest open-source release in Qwen's history. The main line covers seven sizes from 0.5B to 72B, plus specialized lines such as Qwen2.5-Coder and Qwen2.5-Math. Training data was expanded from about 7 trillion to about 18 trillion tokens, and most variants use the Apache 2.0 license.
| Release | Main scales | Key features |
|---|---|---|
| Qwen2 | 0.5B / 1.5B / 7B / 72B, etc. | 128K context, GQA, multilingual |
| Qwen2.5 | 0.5B–72B, Coder / Math series | 18T token training, Apache 2.0 |
| Tongyi Qianwen 2.0 | Proprietary / commercial | Enhanced capabilities for consumer and enterprise |
2025: Qwen3 and Product Ecosystem
Qwen3 MoE architecture and product ecosystem
Technical Milestones
- Qwen3 (April 2025): MoE (mixture of experts) architecture with ~235B total parameters and ~22B active parameters
- Qwen Chat: Alibaba Cloud web experience (chat.qwen.ai) for model demos
- Quark: Positioned as Alibaba's flagship AI app, with a dialogue mode powered by the latest Qwen models
- Qianwen App: Tongyi app rebranded as Qianwen App (v5.0), built on Qwen's strongest model and positioned as a ChatGPT alternative
| Product | Owner | Role |
|---|---|---|
| Qwen Chat | Alibaba Cloud | Web-based model experience and capability showcase |
| Quark | Intelligent Information BU | Flagship AI app: search and dialogue |
| Qianwen App | Intelligent Information BU | Full-scenario AI assistant, ChatGPT alternative |
Summary: Technical Evolution
Qwen has evolved from dense models to MoE (Qwen3), balancing scale and inference cost. It started at 7B and now spans a full range from 0.5B to 72B and a 235B total-parameter flagship. Capabilities have been strengthened in multilingual, code, math, long context, and instruction following, with dedicated lines such as Coder, Math, and VL. Qwen2 and Qwen2.5 are available under Apache 2.0 on Hugging Face and ModelScope, supporting both open and commercial use.
Key Takeaways
- April 2023 launch; Qwen-7B open source in August; Tongyi app in September
- 2024: Qwen2 and Qwen2.5 large-scale open source, 128K context and 18T token training
- 2025: Qwen3 MoE, Qianwen App rebrand, Quark and Qwen Chat form a clear product set
- Dual track: open source and commercial; many variants Apache 2.0 and commercially usable
Try Qwen on FuseAITools for image generation, editing, and multimodal creation.