According to TechNode, Qwen4-Max uses a mixture-of-experts architecture with 2.4 trillion parameters, accepts images directly and supports a context window of up to one million tokens. The model is also available through Qwen Studio, QwenWork and Qoder.
Alibaba said the model uses fewer tokens than Qwen3.8-Max on comparable tasks. Together with lower API prices, the average task cost on the company's Work360 benchmark fell from RMB 26.58 to RMB 11.16. The company also released a plugin that lets agents deliver a first version of their work and revise it based on user feedback.
Alibaba plans to release smaller Qwen4 models as open source, but did not give a date. The cost-saving figures are based on the company's internal benchmark, not independent testing.