Key facts
- Alibaba said Qwen4-Max uses a sparse mixture-of-experts architecture with 2.4 trillion total parameters.
- Alibaba priced Qwen4-Max input at $1.5 and output at $4.5 per million tokens, with cached input at $0.075.
- Alibaba said token use fell about 24% to 49% versus Qwen3.8-Max depending on benchmark and reasoning setting.
- Alibaba said Qwen4-Max scored 52.69 on its Work360 benchmark, against 47.74 for Qwen3.8-Max.
- Alibaba said smaller Qwen4 models will be open-sourced in stages, without dates or variants decided.
Alibaba Group Holding released Qwen4-Max, its most advanced artificial intelligence model, on November 16, saying it cut the average cost of completing a task by about 60% compared with the prior generation. The company said the model uses a sparse mixture-of-experts design, in which only part of the network runs for each request, with 2.4 trillion total parameters.
Alibaba said Qwen4-Max natively handles images and has a context window of 1 million tokens, the units of text a model can process at once. It said the model is available to developers through an API on its Qianwen AI Platform and QwenCloud, and to general users through apps including Qwen Studio, QwenWork and Qoder.
On pricing, Alibaba set input at $1.5 per million tokens and output at $4.5 per million tokens, with cached input at $0.075 per million tokens, 70% below the prior generation. It said the average cost of a Work360 task fell to 11.16 yuan from 26.58 yuan, a decline of 58%.
Alibaba said Qwen4-Max scored 52.69 on its Work360 benchmark against 47.74 for Qwen3.8-Max. It said token consumption at equal performance fell about 24% and 35% at its two higher reasoning settings, and about 48% and 49% on the Terminal-Bench 4.0 test. The Work360 framework covers 26 primary domains, 182 sub-domains and 721 professional directions, according to the company.
Alibaba said it filtered the model's training data to remove offensive cyberattack capabilities while keeping other abilities, and that the release carries cybersecurity guardrails. It said smaller models in the Qwen4 series will be open-sourced in stages, with release dates and variants not yet decided.
