Alibaba Group unveiled its latest reasoning model, Qwen3-Max-Thinking, on Tuesday.

The company said it expanded the model to more than 1 trillion parameters for reinforcement learning. It stressed that this delivered performance improvements across several core areas, including factual knowledge processing, complex reasoning, instruction following, human preference alignment and agent functions.

Alibaba said Qwen3-Max-Thinking showed competitive performance across 19 major benchmark evaluations, compared with the latest high-performance models including Claude Opus 4.5, Gemini 3 Pro and GPT-5.2-Thinking-xhigh. It also posted high-level results in evaluation items that solve questions at an expert level across various fields by using search tools, as well as in solving science, mathematics and coding problems.

The company highlighted two technology innovations related to Qwen3-Max-Thinking.

The first is adaptive tool-use, which supports more efficient problem-solving by allowing the model to search for information depending on the situation and automatically call and use its built-in code interpreter when needed, so users do not have to select tools themselves.

The second is an advanced test-time scaling technique. The company said this improved reasoning performance and delivered results in major reasoning benchmarks that exceeded other high-performance models.