Alibaba unveils new Qwen3 series of open-source AI models

On 29 April, Chinese e-commerce and tech giant Alibaba released its latest generation of large language models (LLMS), the Qwen3, a family of 8 open-source AI models. They have been hailed as the first “hybrid” reasoning models from China, meaning they can combine “flash” fast reasoning with “deep” slow reasoning to solve a single problem, saving on computing power. Alibaba claims that it greatly reduces the cost of deploying these models.

The Alibaba Qwen3 includes 2 MoE (mixture of experts) AI models and 6 dense models. Its flagship model Qwen3-235B-A22B, one of the two MoE models, only has 235B parameters in total, a third of that of DeepSeek-R1. The resources needed for this flagship model are between 25% to 35% of those of the DeepSeek-R1. It only requires a third of the Video RAM (or VRAM, the memory on GPUS) of other models with similar capabilities. In comparison to similar models, it is more powerful than DeepSeek-R1 and OpenAI o1 after multiple tests.

On Weibo, China’s Twitter equivalent, the topic “Alibaba Qwen3 tops global best open-source LLM list” (#阿里千问3登顶全球最强开源模型#) reached number 9 on the Hot Search list with 4.60 million views. Tech and Alibaba-related stocks also saw a surge in Hong Kong.

The competition in the LLM space, especially between the United States and China, is intensifying in 2025 with the “catfish effect” from DeepSeek and the geopolitical tension surrounding tech and chip manufacturing. Since January this year, the 10 major AI companies from the two countries have released 14 base LLMs, including DeepSeek-R1, Alibaba’s Qwen2.5-Max, Google’s Gemini 2.0 and 2.5 Pro, Tencent’s Hunyuan T1, Meta’s Llama 4, ByteDance’s Doubao 1.5, OpenAi’s GPT-4.5, o3 and o4-mini, etc. Some pundits believe the timing of Qwen3’s launch acts as a head start against DeepSeek-R2, rumoured to be released in May. The release is bound to be closely watched by competitors and users alike.


Share

Join our newsletter