The Qwen team launched the inference model Qwen3-Max-Thinking and opened the "adaptive reasoning" experience in Qwen Chat. According to official and related reports, the model improves reasoning, knowledge, tool use and agent capabilities through large-scale training and reinforcement learning, and can automatically decide whether to call tools such as search, memory and code interpreter in dialogue to reduce manual selection and process switching.
The technical selling point also includes "scale-on-test", that is, multiple rounds of self-reflection and correction on complex questions to improve the quality of the final answer. The publicly disclosed benchmark score shows that it scored 98.0 on HMMT Feb and 49.8 in the "Agentic Search" scenario of Humanity's Last Exam (HLE) with search. On the developer side, the Alibaba Cloud Model Studio documentation shows that you can call a specified version of the model through the OpenAI-compatible Chat Completions and Responses interfaces.
FAQs
Q: What problems does Qwen3-Max-Thinking mainly solve?
A: It is aimed at complex reasoning and agency tasks, emphasizing the automatic use of tools to complete information and calculations when needed, and uses multiple rounds of reflection to improve the stability of problem solving.
Q: What does Qwen3-Max-Thinking's "Adaptive Tool Call" mean?
A: The model can automatically select a combination of capabilities, such as search, memory, and code interpreter, in a single task, rather than manually specifying the process by the user.
Q: How does Qwen3-Max-Thinking access through API?
A: You can use OpenAI-compatible Chat Completions or Responses APIs in Alibaba Cloud Model Studio to call the corresponding version according to the document parameters and model name.
