Qwen3-Max

What it is

Qwen3-Max is a proprietary large language model from Alibaba, released in September 2025 and made available through the company's API rather than as downloadable model files ^[1]. At release it had more than 1 trillion parameters and had been trained on about 36 trillion tokens ^[1]. It was Alibaba's largest model at the time ^[1].

Two terms carry most of the weight in that description. Parameters are the numbers inside the network that training adjusts; a model with more of them can in principle represent more intricate patterns, at the cost of more computation to train and to serve. Tokens are the pieces of text — words, parts of words, punctuation — that a model reads and predicts one at a time, so the training-token count describes how much text it was exposed to. A trillion parameters and 36 trillion training tokens both put Qwen3-Max in the largest tier of models in circulation.

"Proprietary" describes how the model is distributed, not how well it works. The weights are not published; users reach the model as a service, sending requests to Alibaba and receiving answers, and the company controls the hardware, the versions and the terms of access.

Where it came from

Alibaba's Max tier is the company's closed flagship line, as opposed to the Qwen models released openly. Qwen3-Max followed Qwen2.5-Max as the new Max-tier flagship, taking over the top position in that line ^[1].

The name first appeared in public as Qwen3-Max-Preview, which The Decoder reported on 7 September 2025 as having more than one trillion parameters ^[2]. At that stage the preview could be used through Qwen Chat, Alibaba's chat product, and through the Alibaba Cloud API ^[2]. A "preview" label generally signals a model that is publicly usable while the final version is still being finished.

Reported performance

The Decoder reported that Qwen3-Max-Preview outperformed Qwen3-235B-A22B-2507, an earlier Qwen3 model, both in Alibaba's internal benchmarks and in the hands of early users; the article did not publish exact scores ^[2]. Internal benchmarks are tests written and run by the company that built the model, so they are a first indication rather than an independent measurement, and early-user impressions are informal by nature.

Alongside the open releases

Qwen3-Max arrived in the same month as the open Qwen3-Next, Qwen3-Omni and Qwen3-VL models ^[1]. The pairing is worth noting because the two kinds of release allow different things: a closed model is reached only through the provider's service, while an open model is published for others to download and run themselves. Alibaba was doing both at once in September 2025.

What came after

Qwen3-Max was followed by Qwen3.5-Plus in February 2026, Qwen3.6-Plus in April 2026, Qwen3.7-Max in May 2026 and Qwen3.8-Max in August 2026 ^[1]. The sequence shows a steady cadence of releases over the following year, with the Max label returning on the later flagships and a Plus label appearing in between. For readers tracking the line, Qwen3-Max is the point where Alibaba's proprietary flagship moved past the trillion-parameter mark and became the reference against which its own later models were presented.