Alibaba has introduced a new AI model named Qwen 3.8 Max, part of the Qwen 3.5 architecture family.
The model has 2.4 trillion total parameters, though it activates only around 95 billion of them for any given task, a design intended to cut computing costs.
Sixteen days, one instruction
In a demonstration of long-horizon autonomous coding, Qwen 3.8 Max was given a single instruction and then operated independently for 16 consecutive days.
During that period, the model handled all coding, testing and debugging itself, without human intervention, ultimately producing a functional command-line tool called oh-my-cli.
Beating the humans
Beyond that solo project, Qwen 3.8 Max also entered a live public coding contest against human competitors on Alibaba Cloud's Tianchi platform.
More than 500 human teams took part in the challenge, which focused on multimodal dialogue intent recognition.
Across 45 submissions over 24 hours, the model's accuracy climbed sharply, and Alibaba said its performance outstripped the majority of the human teams entered.
Part of a wider race
The launch comes shortly after Chinese rival Moonshot AI released Kimi K3, a 2.8 trillion parameter model that currently holds the title of China's largest.
Alibaba has positioned Qwen 3.8 Max as a significant advance in autonomous AI agents, large language models and AI coding assistants.
Open-weight versions of the model are due to follow within days, a move that would make it the largest Alibaba model yet released outside a closed API, and one pitched squarely at developers building longer, less supervised AI-driven workflows.
