Alibaba unveiled a new flagship model in its Qwen family. It packs 2.4 trillion parameters, but thanks to its *Mixture-of-Experts* architecture, only about 95 billion get activated per request.
Published benchmarks show Qwen3.8-Max putting up strong numbers across several categories:
— 93% on PaperBench — reproducing scientific research;
— 86.6% on TerminalBench — working with a terminal;
— 81.8% on LVBench — analyzing long-form video;
— 77.8% on MobileWorld — controlling mobile interfaces;
— 73.5% on FrontierSWE — complex software engineering tasks;
— 69% on Vision2Web — building sites from a visual reference;
— 53.4 points on JobBench — completing work tasks in place of a human.
Qwen3.8-Max stands out especially in tasks where the AI isn't just answering a question but has to carry out a sequence of actions on its own: parsing documents, writing and testing code, operating interfaces, or running research.
The model also works with text, images, and video, and its context window reaches 1M tokens — enough to load entire large documents, long recordings, and big projects in one shot.
Alibaba says it will open-source the Qwen3.8-Max weights. Given the model's size, though, running it locally will mostly be feasible for large companies and research teams — for everyone else, API access and cloud services will remain the main way in.
You can try Qwen3.8-Max right now in Qwen Studio.
It looks like Alibaba isn't just chasing "the smartest model" anymore — it's betting on AI that can stick with a task for the long haul and actually see it through.