Chinese tech giant Alibaba has made its latest and most ambitious move in the artificial intelligence arena at its annual Apsara Conference. The company unveiled Qwen3-Max, billed as its most powerful large language model (LLM) to date. This colossal model stands out as a true parameter behemoth, boasting over a trillion parameters, propelling it far beyond current artificial intelligence technologies. The introduction of Qwen3-Max once again underscores China's assertive stance in the global AI race.
One of Qwen3-Max's standout capabilities is emphasized in its superior performance in coding and autonomous agent fields. The model's "Instruct" version achieved an impressive score of 69.6 on SWE-Bench Verified, a leading benchmark platform for solving real-world coding challenges. Furthermore, it garnered a remarkable score of 74.8 on Tau2-Bench, which evaluates agent capabilities and tool-interaction proficiency, surpassing rivals such as Claude Opus 4 and DeepSeek V3.1. These achievements highlight the potential Qwen3-Max offers to both developers and complex systems.
As part of Alibaba Cloud's strategy to scale its AI systems, Qwen3-Max-Base was trained on 36 trillion tokens. The model's stable and efficient training process was made possible through advanced technologies like the Mixture of Experts (MoE) architecture and PAI-FlashMoE. Consequently, Qwen3-Max-Base achieved a 30% increase in training efficiency compared to Qwen2.5-Max-Base. Its ability to handle long context windows is also noteworthy; through the ChunkFlow strategy, it can perform seamless training with a context length of 1 million tokens. These features enable the model to work more effectively with more complex and lengthy texts.
As a forward-looking move, the company is also continuing its work on a version named Qwen3-Max-Thinking, which is still in the final stages of training. It is reported that this version, when augmented with a code interpreter and parallel test-time compute techniques, has achieved a 100% success rate on challenging mathematical reasoning tasks such as AIME 25 and HMMT. While Qwen3-Max-Instruct is currently available for trial via Qwen Chat, API access has also been made available to users through Alibaba Cloud. Alibaba is expected to further increase its influence in the AI ecosystem with this new large language model.