Alibaba Cloud has quietly released the Qwen3-Max-Preview large model, claiming it has over 1 trillion parameters. According to official release data, this version shows improvements over its predecessor Qwen3-235B in dialogue capabilities, task execution, and instruction following.

Currently, it can only be accessed through the Qwen Chat web interface and Alibaba Cloud API as a closed-source service, which has already drawn sarcastic comments from many in the community. Interestingly, there's a 'thinking mode' toggle in the web interface—users discovered that when turned off, it calls Qwen3-Max, but when enabled, it actually calls Qwen3-235B.

Someone on Reddit posted test screenshots: asking the model to list five countries where the third letter is 'A'. In normal mode, it directly outputs the results, while in thinking mode, it shows the reasoning process. This design is quite clever, maintaining the 'reasoning' ritual while shifting computational costs to the smaller model.

The technical team admitted in the comments that the currently released version doesn't have complex reasoning capabilities. As for the 1 trillion parameter figure, it cannot currently be verified for performance nor can it be deployed locally. For developers accustomed to open-source models, this 'black box' release is like being shown sports car keys but only allowed to sit in the back seat.

What's worth noting are the comments section opinions: some believe that going closed-source is a legitimate business strategy, while others point out that open-source models like GLM4.5 have also shown impressive performance in API tests. Currently, the pricing and specific capabilities of Qwen3-Max remain a mystery, with officials only saying the official version will have 'bigger surprises'—a line we've heard too many times in the large model competition.

Web experience address: https://chat.qwen.ai/