Alibaba has unveiled its new flagship AI model Qwen4-Max. The algorithm with 2.4 trillion parameters uses only 71 billion parameters per token during operation, and also supports image processing. In addition, Alibaba has introduced special datasets for additional training and development to reduce risks in the field of chemical, biological, radiological and nuclear weapons.

Image source: qwen.ai
Another feature: strict cybersecurity restrictions. This likely sees Alibaba trying to balance security with its continued commitment to open source amid growing concerns around the world about AI-powered cyberattacks. The focus on security reflects a growing focus on the issue throughout China’s AI industry. Alibaba’s announcement noted that the company intentionally filtered the training data, removing information that could be used for “offensive” cybersecurity activities, but retaining useful defensive capabilities, such as finding vulnerabilities.
Alibaba highlighted Qwen4-Max’s achievements in scientific research. This includes a continuous 48-hour study that discovered and experimentally verified a refrigerant molecule for next-generation semiconductor products, as well as an offline study of Alzheimer’s disease. While the Qwen4-Max will remain a closed model, the developer has confirmed that smaller versions of the Qwen4 with open weights will be available soon.
As already mentioned, Qwen4-Max has 2.4 trillion parameters, just like the previous Qwen3.8-Max. However, the new AI model only activates 71 billion parameters per token, which is 25% less than the previous version of the algorithm. Thus, the algorithm provides high reasoning abilities while reducing inference costs. According to Alibaba, the average cost of completing one task in Work360, an internal benchmark for assessing the performance of models across different tasks, fell by 60%. Qwen4-Max has a context window for 1 million tokens. The cost of cached input is $0.075 per 1 million tokens. The API price is $1.5 for 1 million input tokens and $4.5 for 1 million output tokens.
If you notice an error, select it with the mouse and press CTRL+ENTER.


