Alibaba Releases Qwen4-Max, a 2.4T MoE, and Keeps the Flagship Weights Closed
Alibaba said on November 16 that it released Qwen4-Max, which the company calls the most capable model in the Qwen family. The launch post gives the model 2.4 trillion total parameters, native vision, and a one-million-token context window. It is a sparse mixture-of-experts model that activates 71 billion parameters per token. Alibaba says that is 25 percent fewer activated parameters than Qwen3.8-Max at the same 2.4 trillion scale. The model is on the QwenCloud API. The post says smaller Qwen4 models will ship with open weights, and it does not offer a weight download for Qwen4-Max.
The numbers
- 2.4 trillion total parameters and 71 billion active per token. Both figures are in Alibaba's November 16 post.
- Around 60 percent lower average cost per task on Work360, and 35 percent fewer tokens at High effort, versus Qwen3.8-Max. Alibaba describes Work360 as its own test across 721 fields.
- $1.50 per million input tokens and $4.50 per million output tokens on the QwenCloud model page. Implicit cache hits are listed at $0.075 per million input tokens. The launch post quotes only the cache-hit price.
- Sixth on Code Arena. That rank is Alibaba's. The Next Web says it has not independently verified the scores.
What shipped
Alibaba says Qwen4-Max is available now on QwenCloud through the Token Plan and the API, and inside QwenWork and QwenStudio. The model page lists text, image, and video as inputs, and text as the output. The launch post says the new architecture changes attention, residual connections, and embeddings inside the sparse mixture-of-experts design.
The Next Web reports a limited preview on November 11 and says the flagship ships without open weights. Alibaba's own sentence is narrower. It says open-weight, smaller-scale models from the Qwen4 series will soon be released, and it gives no date. The Next Web also writes that Alibaba's Work360 average fell from $4.43 to $1.67 a task. Those two dollar figures are not in the company post, which stops at around 60 percent.
What we are watching
The cost claim compares Qwen4-Max with Qwen3.8-Max on Alibaba's benchmark. It is not a published comparison against another lab's invoice. The $1.50 and $4.50 rates sit on the product page, not in the blog's only pricing sentence. No outside lab in these sources has republished the Code Arena rank. A weight file for this model is not part of the release.
Quick answers
- What happened with Alibaba?
- Alibaba said on November 16 that it released Qwen4-Max, which the company calls the most capable model in the Qwen family. The launch post gives the model 2.4 trillion total parameters, native vision, and a one-million-token context window. It is a sparse mixture-of-experts model that activates 71 billion parameters per token. Alibaba says that is 25 percent fewer activated parameters than Qwen3.8-Max at the same 2.4 trillion scale. The model is on the QwenCloud API.
- What are the key numbers?
- 2.4 trillion total parameters and 71 billion active per token. Both figures are in Alibaba's November 16 post. Around 60 percent lower average cost per task on Work360, and 35 percent fewer tokens at High effort, versus Qwen3.8-Max. Alibaba describes Work360 as its own test across 721 fields.
- What happens next?
- The cost claim compares Qwen4-Max with Qwen3.8-Max on Alibaba's benchmark. It is not a published comparison against another lab's invoice. The $1.50 and $4.50 rates sit on the product page, not in the blog's only pricing sentence. No outside lab in these sources has republished the Code Arena rank. A weight file for this model is not part of the release.
This article was ultrathought.