Alibaba says it has released Qwen4-Max, the first model in its Qwen4 family. In its release post, the Qwen team calls it an "AI teammate" for work, creativity and discovery, and says it is a significant step up from Qwen3.8-Max.
What Qwen4-Max is
Alibaba says Qwen4-Max uses a new Qwen4 architecture with a 2.4-trillion-parameter backbone, native vision and a 1M-token context window. It describes the model as a sparse mixture-of-experts design that activates 71B parameters per token, plus an additional 328B N-gram Embedding parameters.
What Alibaba says changed
All figures in this section come from Alibaba's own post and have not been independently verified.
- On Work360, Alibaba says Qwen4-Max scores 52.69, against 47.74 for Qwen3.8-Max.
- Alibaba says token use on Work360 is down about 24% at its Xhigh setting and about 35% at High.
- On Terminal-Bench 4.0, it says token use is down about 48% and 49% at the same two settings.
Alibaba's comparison chart sets Qwen4-Max against Qwen3.8-Max-0902, Opus5.5, GPT6.1-Sol, GPT6-Astra and Kimi K3.1. Alibaba says some of the scores come from third-party leaderboards and others from its own runs.
What an independent test says
Artificial Analysis, an independent benchmarking firm, scores Qwen4 Max 43 on its Intelligence Index (v5.0.1, nine evaluations). That ranks it 38th of 81 models and above the median of 41 for comparable models. On the same chart it lists Kimi K3.1 Pro at 46, plus GPT-6.1 Sol and Opus 5.5, both at max settings, at 55 and 59.
It also says the model generated 127M tokens to complete the Index, which it calls "very verbose" against a 78.6M median, and that running the Index costs an average of about $3.27 per task.
That is a different benchmark from Alibaba's Work360, so it is not comparable with Alibaba's $1.67 per task. Artificial Analysis lists a 984k-token context window, slightly under the 1M Alibaba states. It labels the model proprietary and lists Alibaba Cloud as the provider.
Price and availability
Alibaba says it has lowered API prices compared with Qwen3.8-Max, with particularly large reductions in cached-input pricing. It says the average cost per task on Work360 falls from $4.43 to $1.67, a 62% cut. Alibaba's page lists no per-token prices. Artificial Analysis lists $1.50 per 1M input tokens and $4.50 per 1M output tokens, with a 95% cache discount.
Alibaba says the model is now available on QwenCloud, where you can start building with either the Token Plan or the API. It also lists QwenWork, Qoder and QwenStudio as places to use it.
Open weights
Alibaba says Max is only the first look at Qwen4 and that smaller models in the family are coming to open source. The post does not say when, and it does not say whether Max itself will get open weights.
What we do not know
- Whether Alibaba's own benchmark results hold up. Artificial Analysis's Index is one independent score; we checked no other independent tests.
- Whether weights for Qwen4-Max will ever be released.
- When the smaller Qwen4 models will arrive.
- Which regions can use the model.
Related: GPT-6.1 Terra pricing and benchmarks and Meituan's LongCat 2.5 open weights under MIT.
What we checked
We read Alibaba's post on the evening of Nov. 15 ET; its page is dated Nov. 16. We read Artificial Analysis's model page the same night, after midnight ET. Alibaba's post is our only source for its claims. Artificial Analysis is our only source for the independent score. We have not run the model or the benchmarks ourselves.



