The list of enormous Chinese AI models grew again this week. At the World AI Conference in Shanghai, Alibaba previewed Qwen3.8, a 2.4-trillion-parameter model it is positioning as one of the most capable in the world, with a free, downloadable version promised soon. It landed two days after Moonshot's Kimi K3, which means the two biggest open models on the planet were announced in the same week.
What Alibaba actually showed
Qwen3.8 is not fully out yet. What exists right now is a preview, Qwen3.8-Max-Preview, running through Alibaba's own cloud services at a fraction of its usual price. The full open-weight version, the one anyone can download and run on their own machines, is described only as "coming soon."
Alibaba also calls it the company's first trillion-parameter model that handles images and video as well as text, according to Qwen developer Shuai Bai. So this is a preview and a promise, not something you can pull down and use today.
What "2.4 trillion parameters" means
Parameters are the internal dials a model adjusts as it learns from data. A rough rule of thumb: the more parameters, the more raw capacity a model has, a bit like a bigger engine having more power on paper.
Qwen3.8 has 2.4 trillion of them. Kimi K3, announced two days earlier, has 2.8 trillion. On size alone, you might expect Kimi to come out a little ahead.
But parameter count has never been a clean predictor of real-world quality. Plenty of smaller models beat larger ones on the tasks people actually care about. Until independent testers put these two head to head, the size comparison is just arithmetic.
The claim, and the missing proof
Alibaba says Qwen3.8 trails only Anthropic's Fable 5 among the world's models. That is a bold thing to say, and for now it is only a claim: the company has not published any test results, and no neutral party has checked the boast.
Reactions online split about how you would expect. Some people are excited to have another powerful open model on the way; others are asking to see real numbers before they believe the ranking. The reasonable posture for the rest of us is interest without belief until the tests actually show up.
This is another big open model, with DeepSeek's V4 rumored to land this month too. The open-source flood keeps putting real pressure on the expensive closed models, which, as I have said throughout this issue, is good news for anyone paying for AI.
I'm excited to put Qwen3.8 through its paces once the open weights actually drop.