Qwen3 Max

Qwen3-Max is a large-scale language model from the Qwen family, positioned as the most capable release in the series so far. The base model was pretrained on 36 trillion tokens and crosses the one-trillion-parameter mark, using the same MoE mixture-of-experts design paradigm as the broader Qwen3 series, which the team credits with keeping the pretraining loss curve smooth and stable. This architectural lineage and scale shape its role as a general-purpose flagship tuned for demanding knowledge, reasoning, and multilingual tasks rather than a narrow specialist.

In its instruct-tuned form, Qwen3-Max-Instruct has been benchmarked across a broad suite covering knowledge, reasoning, coding, instruction following, human preference alignment, agent tasks, and multilingual understanding, with the official release described as achieving state-of-the-art results across that comprehensive suite. A separate thinking-oriented variant, Qwen3-Max-Thinking, targets deeper reasoning through scaled test-time compute and adaptive tool use, with reported benchmark results on tasks such as AIME 25 and HMMT. Together these variants make Qwen3-Max a strong practical fit for complex analytical workflows, agent-style applications, and long-context tasks where the model can lean on retrieval and code execution.