In early November, the Qwen team released an early preview version of Qwen3-Max-Thinking, stating that the model was an intermediate checkpoint still under training. The official statement indicated that, after combining tool usage with expanded test-time compute, the model achieved 100% scores on challenging inference benchmarks such as AIME 2025 and HMMT. The current version is available on Qwen Chat and can be accessed via the Alibaba Cloud Model Studio API by enabling the enable_thinking parameter.
It's important to note that publicly available third-party leaderboards typically use fixed settings and may not account for computational power expansion during external tools or unconventional testing. Therefore, their results may differ from those claimed by manufacturers as "tool enhancements + expanded computational power." Recent AIME 2025 summary leaderboards do not generally display 100% perfect scores; whether they will be included in future unified rankings depends on the evaluation rules and reproduction procedures. Overall, this release is a feature preview; training and metrics will continue to be updated.
Frequently Asked Questions
Q: Where can I use Qwen3-Max-Thinking now?
A: You can try it out in the Qwen Chat frontend, or you can call it through the Alibaba Cloud Model Studio API and set enable_thinking=True in the request to enable thinking mode.
Q: What are the specific conditions for the claimed AIME 2025 and HMMT "100%"?
A: The official explanation is that it was obtained under the conditions of "enhanced tools + expanded inference computing power during testing"; there is a difference in the definition compared to the public leaderboard with standard closed settings.
Q: Why do public rankings not necessarily show perfect scores?
A: Many rankings require a fixed temperature, no external tools, or a limited inference budget; scores may differ or not be included if the test setup differs from the official test setup.
Q: Is this the official version?
A: No. This version is an early preview and is still under development. Its capabilities and stability may change in the future. The official statement is that it will continue to be updated.
Q: How do I enable the thinking mode in the API?
A: Use the enable_thinking parameter in the relevant interfaces of Alibaba Cloud Model Studio; the specific implementation documentation provides examples.
