AlphaGazette
Menu

AI

Alibaba launches closed-weight Qwen4-Max with 2.4 trillion parameters

The first Qwen4 model keeps 71 billion parameters active per token, reads a one-million-token window, and starts at $1.50 per million input tokens on QwenCloud. Alibaba says it uses up to ~50% fewer tokens than Qwen3.8-Max and plans open weights only for smaller Qwen4 models.

2 min read 6 sources AI-generated

Alibaba on Monday, 16 November launched Qwen4-Max, a reasoning model and the opening release of the Qwen4 family, with the flagship weights kept closed. Developers can use it for coding, professional work and agents that run a long time, through QwenCloud and, under the name alibaba/qwen4-max, on Vercel AI Gateway from the coding agents Claude Code, Codex and Cursor. It is also offered in QwenWork, in Qoder and in QwenStudio.

Architecture and price

Alibaba said the system is a new mixture-of-experts design: 2.4 trillion parameters in the backbone, 71 billion active on each token, and 328 billion further parameters for N-gram Embedding, which is meant to handle patterns in text more efficiently. It accepts images and text and has a one-million-token context window. Input on QwenCloud is priced from $1.50 per million tokens, and output from $4.50 per million tokens. Vercel said the gateway bills the provider's own rate, adds nothing on top, and takes no fee for inference, even on bring-your-own-key traffic.

Benchmarks and later releases

Alibaba said Qwen4-Max beats Qwen3.8-Max and uses up to ~50% fewer tokens, with the gains showing on complex work that calls many tools and continues on its own for a long stretch. On Work360, its internal score for professional tasks across hundreds of specialised fields, the new model reached 52.69 and Qwen3.8-Max reached 47.74. Alibaba says coding benchmarks put Qwen4-Max in front of Fable 5.1 and of Kimi K3.1 Pro. The Decoder said the model trails Kimi K3.1 Pro by a small margin on the Artificial Analysis Intelligence Index and that a hybrid design cuts the bill for long contexts.

The Qwen team pitched the release as an "AI teammate for work, creativity and discovery." Alibaba had teased the family last week and first made Qwen4-Max a preview in several of its products. It plans an open-weight release of the smaller Qwen4 models. For long tasks, the flagship is built to deliver a usable early version and then keep improving it, and Alibaba is releasing a Live Delivery plugin as open source for that workflow.