Alibaba launches Qwen3-Max-Thinking reasoning model
Alibaba's Qwen team has released Qwen3-Max-Thinking, a flagship reasoning model it says matches GPT-5.2-Thinking, Claude Opus 4.5 and Gemini 3 Pro on 19 benchmarks, with tools it decides to call on its own.
Alibaba's Qwen team announced Qwen3-Max-Thinking, its new flagship reasoning model. The team says it got the gains by scaling up model parameters and spending heavily on reinforcement learning compute.
In the team's own evaluations across 19 benchmarks, the results are mixed against frontier rivals. Qwen3-Max-Thinking leads on HLE with tools (49.8, against 45.8 for Gemini 3 Pro), Arena-Hard v2 (90.2, judged by GPT-4.1), HMMT Nov 25 (94.7) and C-Eval (93.7). It trails on SWE-bench Verified (75.3, against Claude Opus 4.5's 80.9), MMLU-Pro (85.7), GPQA (87.4) and Vita Bench (40.9). None of these scores have been independently verified.
The model also chooses for itself when to use its built-in Search, Memory and Code Interpreter tools, so users no longer pick tools before each task. Qwen says it trained this behavior with tool-use fine-tuning followed by further training on rule-based and model-based feedback.
A separate "heavy mode" uses what Qwen calls experience-cumulative, multi-round test-time scaling. Rather than adding more parallel reasoning paths, it runs fewer of them and uses a "take-experience" step to carry key insights from one round of self-reflection into the next. The team says this beats standard parallel sampling at roughly the same token cost, raising GPQA from 90.3 to 92.8 and LiveCodeBench v6 from 88.0 to 91.4.
The model is live in Qwen Chat and through Alibaba Cloud Model Studio as qwen3-max-2026-01-23. The API works with both the OpenAI and Anthropic API formats, which lets it run inside Claude Code. The post does not give pricing, parameter count or weight availability.