Alibaba launches the Tongyi Qianwen chatbot model

Background

Tongyi Qianwen (通义千问) was the large language model project of Alibaba, the Chinese technology company. A large language model is a program trained on very large amounts of text so that it can read and write natural language. According to Wikipedia, the original architecture of Tongyi Qianwen was based on Meta's Llama, a family of models released by another company.^[1] The project later took the international brand name Qwen.^[1]

What happened

Alibaba launched Tongyi Qianwen as a beta in April 2023.^[1] A beta is an early version offered for testing before a full public release. The model was opened to the public in September 2023, after regulatory clearance.^[1]

Between the beta and the public opening, the first open-weight model in the family, Qwen-7B, came out on 3 August 2023.^[2] Open weights are the trained parameters of a model, published so that others can download it and run it on their own hardware.

Key details

The beta and the public opening were separate steps, about five months apart. The public opening came after regulatory clearance, which means the service was made available once the required approval had been given.^[1]

The open-weight releases continued later in the year. Qwen-72B and Qwen-1.8B were made available for download on 30 November 2023.^[2] The sequence therefore shows two paths in the same project: a model offered to the public as a service, and models whose weights were published for download. The two paths ran side by side during 2023.^[1][2]

Impact

Over time the Tongyi Qianwen name gave way to Qwen as the international brand, which is how the family is now generally known.^[1] For a reader meeting these names for the first time, the practical point is that Tongyi Qianwen and Qwen refer to the same family of models. Tongyi Qianwen is the name used in China, and Qwen is the international brand.^[1]

The entries that follow in this series describe the individual models released under the family's names, starting with the first open-weight model and continuing through later generations.