NEWS → · Model launch · Alibaba Qwen
Qwen4-Max is live: a 2.4-trillion-parameter flagship at $1.50 per million input tokens — API only, with open weights promised for the smaller models
Nov 16, 2026
Alibaba's Qwen team has made Qwen4-Max generally available: a 2.4-trillion-parameter mixture-of-experts flagship with native vision and a one-million-token context, priced on QwenCloud at $1.50 per million input tokens and $4.50 per million output — 25% below its own predecessor. There are no weights to rent a GPU for: Max is API-only, and the open-source release Qwen promises is for the smaller models in the family, with no names or dates yet.

Image: Alibaba Qwen via qwen.ai
What shipped, from Qwen's own launch post ('Qwen4-Max: Your New Teammate Has Arrived', qwen.ai, read first-hand on Nov 16, 2026) and Reuters' same-day report: Qwen4-Max is built on the new Qwen4 architecture — a sparse mixture-of-experts model with a 2.4-trillion-parameter backbone, 71 billion parameters activated per token, plus 328 billion N-gram embedding parameters. It takes text, image and video input natively, works over a one-million-token context window, and can return up to 393,216 output tokens per QwenCloud's model page. Access is through the QwenCloud API (model ID qwen4-max, live Nov 16 after a Token-Plan-only preview that opened Nov 11), Qwen Studio and Qwen's own apps. The preview ID, qwen4-max-preview, is a separate thing: it stays on the Token Plan until it is retired on Nov 25 — the generally available model is not affected.
The price, from QwenCloud's published rates checked Nov 16, 2026: $1.50 per million input tokens, $4.50 per million output tokens and $0.075 per million cached input tokens — against $2.00 / $6.00 / $0.25 for Qwen3.8-Max, a 25% cut on input and output and 70% on cached input. Qwen's own efficiency figures (the maker's numbers, not independently rerun at launch) say the model also uses fewer tokens per task: on its Work360 suite the average cost per task falls from $4.43 to $1.67 while the score rises from 47.74 to 52.69. On Alibaba's own 16-benchmark chart the model leads on four (including Agents' Last Exam and DeepSWE) and trails rivals clearly on others, including ProgramBench and Terminal-Bench — read the chart as the maker's case, not a verdict.
What it means if you rent GPUs
For GPU renters the honest headline is the one in Qwen's closing paragraph: 'Max is only the first look at Qwen4. Smaller models in the family are coming to open source.' No smaller Qwen4 weight had been released as of Nov 16 — there is no Qwen4 repository in Qwen's Hugging Face organisation — so today this launch changes API maths, not rental maths: a hosted flagship got 25% cheaper per token, and every self-hosting comparison now has a lower API bar to beat. The rental story starts when a smaller Qwen4 weight actually lands with a licence and a parameter count we can do memory maths on; our model catalogue will carry it then, with the GPU arithmetic shown, and not before. Until then, treat any 'deploy Qwen4-Max on your own GPUs' offer with suspicion — there are no official weights to deploy.