Qwen3 Max

Qwen3 Max is a flagship scale large language model in the Qwen3 series, officially launched with over one trillion parameters and pre-trained on 36 trillion tokens of data. As an updated release, it builds on a January 2025 version with substantial gains in reasoning, instruction following, multilingual coverage, and long-tail knowledge. The model is positioned for high-accuracy work across mathematics, coding, logic, and science tasks, while delivering stronger performance on open-ended question answering, writing, and conversational scenarios in both Chinese and English. It also targets complex workflows by supporting over 100 languages, more reliable translation, and commonsense reasoning, making it suitable for organizations that need a general-purpose model with broad linguistic reach.

The model is explicitly optimized for retrieval-augmented generation and tool calling, allowing it to plug into knowledge pipelines and external systems rather than relying solely on its own parametric memory. Notably, it does not include a dedicated thinking mode, framing it as a unified conversational and agentic model rather than a specialized reasoning variant. Independent benchmarking in the launch coverage reported a SWE-Bench Verified score of 69.6, indicating strong software-engineering and agent capability, alongside a third-place ranking on the LMArena text leaderboard. A companion thinking-focused variant separately achieved 100% accuracy on AIME25 and HMMT for mathematical reasoning, though that mode is delivered through a sibling model rather than Qwen3 Max itself. Practically, this combination makes Qwen3 Max a strong fit for enterprises seeking a large, multilingual, tool-aware model that can handle both analytical tasks and creative or conversational workloads within a single deployment.