Alibaba releases Qwen3-Max, a model of over 1 trillion parameters

Background

Alibaba's Qwen family has run on two tracks. Some models are published as open models whose weights can be downloaded and run by others; others are proprietary, meaning the weights stay inside the company and outside developers reach them through an API — an interface where a request is sent to Alibaba's servers and a response comes back. The two tracks serve different purposes: open models spread the family's name and let researchers inspect and adapt it, while proprietary models are the ones the company serves directly to customers.

The release

In September 2025 Alibaba released Qwen3-Max as a proprietary model available through the company's API^[1]. It has more than 1 trillion parameters and was trained on about 36 trillion tokens^[1]. At release it was Alibaba's largest model, and it followed Qwen2.5-Max as the flagship of the proprietary Max tier^[1].

The Decoder reported on 7 September 2025 that Alibaba unveiled Qwen3-Max-Preview as a model with more than one trillion parameters, and that it was available through Qwen Chat as well as the Alibaba Cloud API^[2]. The preview name and the chat access point suggest a staged rollout: a version exposed to users and developers first, rather than a model announced only on paper.

What the numbers describe

A parameter is one of the numbers inside a neural network that training adjusts. Large language models learn by reading text and repeatedly nudging those numbers so that the model gets better at predicting what comes next; the parameter count is simply how many such adjustable numbers the network contains. Parameter count is a rough measure of scale, not a measure of quality — a bigger model is not automatically a better one, and the figure says nothing on its own about how well a model reasons, how fast it runs, or what it costs to use.

A token is the unit that text is chopped into before the model sees it — sometimes a whole short word, sometimes a fragment of one. Training on about 36 trillion tokens describes the total volume of text the model was exposed to during training^[1]. That figure matters because scale in training data and scale in parameters are usually discussed together: a very large network needs a very large amount of text to be trained on.

Two tracks in the same month

The release did not happen in isolation. Qwen3-Next, Qwen3-Omni and Qwen3-VL — open models — came out in the same month^[1]. So in a single month Alibaba shipped both open models, which anyone can take away and run, and a closed trillion-parameter flagship reachable only through its own services. That combination is a common pattern among large AI labs: open releases build adoption and goodwill, while the closed flagship is the product that customers pay to call.

Why it drew attention

The headline number in coverage was the trillion-parameter scale, reported as the defining feature of the preview^[2]. Within Alibaba's own lineup, the significance is easier to state: Qwen3-Max marked the top of the proprietary Max tier, taking over from Qwen2.5-Max, and it was the largest model the company had released at that point^[1]. For developers already building on Qwen through the API, it was the new upper bound of what the family offered; for everyone else, it was a marker of how large the models behind a commercial API had become.