Alibaba officially released its new-generation foundation model Qwen4 on November 16. The Qwen AI platform launched Qwen4-Max first, with API services and the Token Plan opening simultaneously. The model adopts a new MoE architecture based on Qwen Sparse Attention (QSA), with 2.4T total parameters, native vision and stable delivery for long-horizon tasks, support for million-token context, and the ability to call specialized tools to handle long-cycle tasks and deliver results in stages based on user feedback.

In the Terminal-Bench4.0 evaluation, at the same Xhigh level, Qwen4-Max reduced average Token usage per attempt from 289K to 151K compared with the previous generation Qwen3.8-Max, a decrease of nearly 50%; its score rose from 37.9 to 48.5, an increase of nearly 28%; cache price cuts drove a 58% reduction in average cost per task.
In terms of pricing, domestically the cost per million Tokens is 10 yuan for input, 30 yuan for output, and 0.5 yuan for implicit cache hits; the corresponding international site prices are US$1.5, US$4.5, and US$0.075. In addition to the Qwen AI platform, it can also be accessed via QwenCloud or experienced in applications such as Qwen Office and Qoder.
In terms of deployment, the model integrates drawing understanding, parametric modeling, and tool calling into the automotive manufacturing process, reducing manual input for single-workpiece machining programming from about 475 minutes to 26 minutes, a reduction of about 95%; in the AI4S field, after 48 hours of continuous agent operation, it built a molecular candidate space of 2.7 million molecules and obtained experimentally validated new coolant molecules through 4 closed-loop rounds. Around the NLRP3 target, it cumulatively called tools 454 times, completed 3,239 metric evaluations, and optimized 24 candidate molecules, with the nominal binding affinity of the selected molecule reaching about 3.6 times the patent starting point.
In addition, the model also participated in optimizing its own scheduling, reducing data annotation scheduling costs by 9.00% versus baseline, increasing the duration of idle capacity in GPU online allocation by 4.62%, and improving CPU offline rescheduling solve speed by more than 14.61 times. From tool calling to phased delivery, this launch continues the trend of large model competition shifting from capability validation to real productivity deployment.





