Qwen4-Max is the first model on Alibaba's Qwen4 architecture and its current flagship. It is a sparse mixture of experts with a 2.4T-parameter backbone, 71B active per token, plus 328B N-gram embedding parameters. Context is 1M tokens, with a maximum input of 991K, or 983K with thinking enabled, and reasoning able to consume up to 262K. Inputs are text, image and video; output is text only.
Alibaba describes it as a long-horizon agent model rather than a chat model, and its own launch chart supports that. It leads Agents' Last Exam v1.1 at 46.7 against 43.4 for Claude Opus 5.5 and GPT-6 Astra, and the HLAB-Vals-1.1 legal-agent benchmark at 21.3 against 11.7 for Kimi K3.1 and 9.2 for Opus 5.5 - roughly double the nearest rival.
Coding is the weak spot, on Alibaba's own numbers. Terminal-Bench 4.0 comes in at 48.5, behind Kimi K3.1 at 67.2, Opus 5.5 at 59.6, GPT-6 Astra at 59.1 and GPT-6.1 Sol at 56.1. ProgramBench is 31.5 against 65.0 for Opus 5.5. It wins DeepSWE v1.1 at 75.7, but by half a point over GPT-6.1 Sol, which is noise. If you are driving a coding agent, the saving comes with a real capability cost.
Independent placement runs lower than the launch chart. Artificial Analysis scores it 43 on its Intelligence Index, ninth overall. Vals ranks it 11th of 38 at 63.20%, and its own Terminal-Bench run of 48.48% matched Alibaba's published 48.5 almost exactly - so the figures appear honest, the model is simply mid-table. Testers consistently describe it as slow and verbose.
Pricing on QwenCloud's international list is $1.50 per million input tokens and $4.50 output. Cached input and explicit cache reads are $0.075 per million; cache creation is $1.875 on the five-minute window or $3.00 on the one-hour. There is a free quota of 1M tokens in the Singapore region for around 90 days, and a Token Plan subscription starting around $6 a month on third-party reports.
Access is through the qwen4-max model ID on QwenCloud, with both OpenAI-compatible and Anthropic-compatible endpoints, so swapping it into an existing stack is a config change rather than a rewrite. It also appears in QwenStudio, QwenWork and the Qoder coding product.
Two things in the documentation do not line up. The output limit is given as 131,072 tokens without thinking and 262,144 with it in the API docs, and as 393K in both modes on the model page. And Alibaba reports average cost per Work360 task falling from $4.43 to $1.67, about 62%, which is a larger drop than its own token and price reductions account for - unexplained in the launch material.
Weights are closed and there is no release planned for Max. Alibaba says smaller Qwen4 models will be open-sourced later, which follows the previous generation's pattern.