Qwen3-Max is positioned as the largest and most capable model in the Qwen family, scaling the Qwen3 design paradigm with a global-batch load balancing loss for training stability. The base model carries over one trillion parameters pretrained on 36 trillion tokens, building on the MoE Mixture of Experts architecture used across the Qwen3 series. This lineage emphasis on stability and scale reflects an intent to push pretraining further while keeping the loss curve smooth, rather than introducing a fundamentally new architecture.

For practical use, the Instruct variant is delivered as a text-in/text-out model accessible through Alibaba Cloud API and Qwen Chat, with the preview already ranking third on the Text Arena leaderboard ahead of GPT-5-Chat at announcement. The official release targets gains in coding and agent workflows, with state-of-the-art results claimed across benchmarks covering knowledge, reasoning, coding, instruction following, human preference alignment, agent tasks, and multilingual understanding. A companion Thinking variant, augmented with tool usage and scaled test-time compute, has reportedly reached perfect scores on AIME 25 and HMMT, signaling strong potential for math and competition-style reasoning once publicly released.