Alibaba’s Qwen team has announced an early preview of a new model, Qwen3-Max-Thinking, noting that it is an intermediate checkpoint from an ongoing training process.

Despite being a preview version, the team reports that when this model is equipped with external tools and allowed significant computational time for problem-solving, it has achieved a 100% score on recent, challenging high-school and university-level math competitions. Specifically, the model was tested on the 2025 American Invitational Mathematics Examination (AIME) and the Harvard-MIT Mathematics Tournament (HMMT), benchmarks known for their difficulty.

This "thinking" mode allows the model to spend more computational effort to work through complex, multi-step reasoning problems. The current version is available for public testing, offering a look at its developing capabilities.

Availability

Users can try Qwen3-Max-Thinking now through the following channels:

1. Qwen Chat: Access the model directly on the Qwen Chat website. Be sure to enable the thinking feature. Link: https://chat.qwen.ai/?thinking=true
2. Alibaba Cloud API: Developers can integrate the model via Alibaba Cloud's API by specifying the correct model name and parameter.
3. Console: https://modelstudio.console.alibabacloud.com/
4. Model Name: qwen3-max-prmodel
5. Required Parameter: enable_thinking=True

The Qwen team has stated that this is a preview and that further improvements are expected as the model's training continues. This release provides a tangible look at the current state of progress in developing AI systems capable of advanced logical reasoning.