Alibaba has unveiled Qwen3.8-Max, its largest artificial intelligence model to date. According to the company, the system has 2.4 trillion parameters, supports text, images and video, and can handle contexts of up to one million tokens. It can therefore analyse extensive code repositories, legal documentation or reports running to hundreds of pages in a single go. In terms of scale, it is close to Moonshot AI’s Kimi K3, which has 2.8 trillion parameters. A test version of Qwen3.8 is already available on Alibaba’s services, and the company has announced plans to publish the model’s weights openly.
However, it is not the size itself that is key, but efficiency. Qwen3.8-Max utilises an ‘expert mix’ architecture, which is why it activates around 95 billion parameters for a single task. This is intended to reduce computational costs and response times. Alibaba also claims that the model independently completed a software development project in 16 days. The significance of this result can only be assessed once the methodology has been published and independent tests have been carried out.
The launch comes amid growing price pressure in the AI market. In June, models with open weights accounted for 29 per cent of the tokens processed by Vercel’s AI gateway, but for less than 4 per cent of expenditure. For businesses, this means greater choice and the ability to route simpler tasks to cheaper systems. Open-source distribution also facilitates independent hosting, although such a large model still requires costly infrastructure.
For Alibaba, Qwen is a tool for driving cloud sales. In the quarter ending in March, Alibaba Cloud’s external revenue rose by 40 per cent, with AI products already accounting for 30 per cent of those sales. Qwen 3.8-Max may reinforce this trend, but at the same time it will accelerate the price war, increase demand for data centres and intensify the technological rivalry between China and the US.
