According to the source, Apple decided to partner with Alibaba because it saw the strength of Qwen AI model and the company's capabilities. The iPhone maker is looking for a local AI partner to boost sales in the world's largest smartphone market because Apple Intelligence is not available there.

Apple plans to hold its China developer conference in Shanghai on March 25, sparking speculation that Apple Intelligence will appear on that day.

alibaba qwen
Alibaba unveils more than 100 open-source AI models of varying sizes. Photo: Shutterstock

According to the Hugging Face machine learning platform and community, Alibaba's Qwen AI model is behind the top 10 open-source large language models (LLMs).

Hugging Face’s latest rankings show that at least seven of the world’s top 10 most highly rated models were trained and developed on upgraded open-source versions of Qwen. Qwen is part of the Tongyi Qianmen LLM “family” built by Alibaba.

The new inference model from US computer scientists, including AI godfather Li Feifei, was trained for less than $50 on the Qwen2.5 32b-Instruct model, according to research published last week.

Qwen2.5 32b-Instruct was born during the Lunar New Year, quickly climbing to 7th place on Chatbot Arena, a calibration project developed by scientists at the University of California.

Open source allows anyone to modify and share technology thanks to its publicly accessible design. Open source technology has contributed significantly to China's booming technology industry over the past decades.

Qwen's dominance reflects China's growing AI prowess, alongside recent breakthroughs by startup DeepSeek.

DeepSeek made headlines when it announced two AI models—DeepSeek V3 and DeepSeek R1—at a fraction of the cost of its competitors. Meanwhile, according to Alibaba’s latest data, the company’s open-source AI models have led to the creation of more than 90,000 models on Hugging Face.

Alibaba introduced Qwen2.5 in September 2024, announcing more than 100 open source models with parameters ranging from 500 million to 72 billion.

At the time of launch, the Qwen2.5-72b beat several other open source models like Meta's Llama3.1-405b in calibration tests.

(According to SCMP)