Alibaba to open-source Qwen3.8-Max, its biggest AI model yet

On August 3, 2026, Alibaba released Qwen3.8-Max, a 2.4 trillion parameter model with 95 billion parameters active per token. In its announcement the Qwen team wrote that this “marks the first time we will open-source the weights of a Qwen-Max-class model”, with the weights shipping next week. Until now the Max tier stayed behind an API.

What Alibaba says the model did

From the company’s own published writeup:

• More than 10 days of autonomous coding on a public repository, reaching 265 commits, 127 pull requests and 151 issues by July 30
• Reproduced a research paper starting from nothing but the PDF in about 125 hours, then invented a method that beat the paper by 2.7 points on the AIME24 math benchmark
• Entered a live Tianchi contest against 526 human teams and finished ahead of 458 of them within 24 hours

Read the numbers with the right label

These are Alibaba’s own results, self-reported. On TerminalBench 2.1 the model scores 86.6, which trails GPT-5.6 Sol at 88.8, and no independent lab has published a full comparison table yet. The market reacted anyway: Alibaba’s Hong Kong shares rose about 6% on the news.

Why the open-weights part is the story

A 2.4 trillion parameter frontier-class model published for anyone to download would be the largest of its kind so far. Open weights at this scale change who gets to run frontier-class AI: not just API customers, but anyone with the hardware, in any country, on their own terms.

If the weights land next week as promised, the question moves from what Alibaba’s model can do to what everyone else does with it.