Alibaba has released Qwen4-Max, a mixture-of-experts model with 2.4 trillion total parameters and 71 billion active for each token. The model accepts images and offers a one-million-token context window, while a hybrid architecture compresses earlier context before selecting the portions that need closer attention.
Alibaba says the design lowers the average cost of a task on its Work360 benchmark by 62%, from $4.43 to $1.67, while increasing the score from 47.74 to 52.69. External evaluator Artificial Analysis measured a cost of $3.27 per Intelligence Index task, about one-third of Qwen3.8-Max, with Qwen4-Max scoring 43 points. It ranked behind Kimi K3.1 Pro at 46.
Performance varies by test. Qwen4-Max led the compared group on Agents’ Last Exam v1.1 but trailed OpenAI models and Claude Opus 5.5 on Humanity’s Last Exam. It scored 48% on Terminal-Bench 4.0, compared with 67% for Kimi K3.1 Pro.
Alibaba charges $1.50 per million input tokens and $4.50 per million output tokens through its API. A “Live Delivery” mode produces an early usable draft during long agent jobs and lets users redirect the task before later passes finish; Alibaba has open-sourced the supporting plugin.