China's tech giant Alibaba on Monday unveiled Qwen3.8-Max, its largest and most capable AI model to date, intensifying the country's race to develop powerful yet cost-efficient artificial intelligence systems.

The model features 2.4 trillion parameters, second only to Moonshot AI's Kimi K3 with 2.8 trillion parameters. It also ranked among the world's leading AI models on Arena, placing fifth on the Text Arena leaderboard and second on the Vision Arena leaderboard. Overall, Alibaba said its performance is in the top tier globally, trailing only Anthropic's Claude family.

Built on a sparse mixture-of-experts architecture, Qwen3.8-Max assigns tasks to specialized components rather than activating the entire model for every query. It uses only 95 billion parameters at a time, significantly reducing costs and latency while handling text, image and video processing with a context window of up to 1 million tokens.

Alibaba said the model delivers "comprehensive improvements across coding, work, research, and long-horizon tasks," enabling it not only to handle more complex queries but also to complete multi-step tasks end-to-end with greater reliability and produce more dependable outputs.

In a demonstration of its autonomous coding capabilities, Alibaba said Qwen3.8-Max completed a multi-day software engineering project from scratch without human intervention, ranking fourth on the Frontend Code Arena leaderboard with a score of 1,668.

Alibaba also opened API access to developers worldwide, pricing the model at about 40% of Claude Opus 5 for input tokens and 24% for output tokens in international markets, while offering substantially lower costs through cache hits.

Alibaba plans to open-source the model's weights next week.

The release follows growing momentum in China's AI sector. In July, Chinese startup Moonshot AI drew global attention with the releasing of Kimi K3, a 2.8-trillion-parameter open-weight model that currently holds the record as the world's largest open-source AI model by parameter count.

(With input from Reuters)