Alibaba has just unveiled Qwen 3.8-Max, an artificial intelligence model capable of programming completely autonomously for over two weeks without human intervention. The Chinese giant has achieved with this AI what seemed like science fiction just months ago: a machine that not only writes code, but improves, debugs and iterates it by itself, accumulating hundreds of changes in real repositories.

An experiment that ran for 16 days non-stop

According to hipertextual.com, Alibaba demonstrated its model's capabilities by letting it work on a real project called Oh My CLI, a terminal tool. The result was spectacular: Qwen 3.8-Max accumulated 265 commits, 127 pull requests and 151 resolved issues over those 16 days without a single engineer touching the keyboard.

In another equally ambitious experiment, the model worked for nearly five days reproducing an academic study on data selection for AI training. The machine wrote 7,600 lines of code and executed 33 training rounds on GPU, even managing to improve the original results with a gain of 2.7 points on the AIME24 mathematics exam.

Sophisticated technique: fewer active parameters, faster speed

The secret behind this capability lies in its architecture. Qwen 3.8-Max has 2.4 trillion total parameters, but uses a technique called Mixture-of-Experts (MoE) that divides the model into specialists. The clever part is that it only activates 95 billion parameters per request, reducing response times without losing computational power.

The model also supports context of up to one million tokens, meaning it can process enormous conversations without losing the thread. And it doesn't just understand text: it handles images and video in the same conversation, making it a truly multimodal tool.

Outperforms Claude and GPT-5.6 in programming

In performance tests published by Alibaba, Qwen 3.8-Max delivers striking results. It outperforms Claude Fable 5 and GPT-5.6 Sol on seven programming benchmarks and leads on 36 vision and multimodal capability tests. On real-world programming benchmarks, where everyday developer tasks are simulated, the Chinese AI reaches 86.6 points on Terminal Bench 2.1, above Claude Fable 5's 84.6.

Where Qwen falls short is on tests evaluating the resolution of real bugs reported by developers. There Claude Fable 5 achieves 80 points versus Qwen's 67.7. However, Qwen excels on complex medical questions and applied legal knowledge, demonstrating that its strengths are not limited to code.

Available now, open source coming very soon

The model is already available through QwenCloud for anyone who wants to test it. But the most disruptive part is that Alibaba promised to release the model weights as open source next week via Hugging Face and ModelScope. This stands in stark contrast to OpenAI and Anthropic's strategy, which keeps their most powerful models closed.

This decision makes U.S. tech giants nervous, as they have seen a wave of competitive Chinese models arrive on the market this year without warning. Will they maintain their lead when an open-source AI with comparable performance is within everyone's reach?