Alibaba’s Qwen team just released Qwen4-Max, its new flagship AI model and the first in the Qwen4 family. Its main network has 2.4 trillion parameters, but it uses only 71 billion of them for each token it reads or writes.
It can also read images and video, and take in up to a million tokens at once. The model isn’t open source, though smaller Qwen4 models will be. In a post on its official blog, the Qwen team said Qwen4-Max is out now on QwenCloud, its platform for developers.
How it reads long text
Besides using only part of the model on each token, Qwen4-Max mixes two kinds of layers to read long text. For every layer that looks back over the whole text, three other layers squeeze the text so far into a fixed-size summary. The full layer uses a small, fast index to pick out only the parts that matter, instead of going over all of it. That cuts the cost of working through long text.
Fewer tokens and lower prices
Alibaba said Qwen4-Max beats Qwen3.8-Max, its last flagship, in nearly every area it tested. It also used fewer tokens to get there, up to 49% fewer on Terminal-Bench 4.0, a coding test. On Work360, Alibaba’s own test of professional work, it used 24% to 35% fewer. Alibaba also said the average cost per task in that test fell 62%, from $4.43 (about Php 283) to $1.67 (about Php 107).
API prices are lower, too, with input at $1.50 (about Php 96) per million tokens, down from $2 (about Php 128). Output is $4.50 (about Php 288), down from $6 (about Php 384). Cached input saw the biggest cut, from $0.25 (about Php 16) to just $0.075 (about Php 4.80) per million tokens.
Live Delivery for long jobs
Live Delivery is how Qwen4-Max handles long jobs. It hands you a usable first version early, then keeps improving it as it goes. You can step in at any point to fix a mistake or change direction. Alibaba is also open-sourcing a plugin that adds Live Delivery to OpenAI’s Codex CLI, with other tools planned.
What an independent test found
Artificial Analysis is an independent firm that tests AI models. Its page for Qwen4-Max shows the model scored 43 on the firm’s Intelligence Index. That’s above the median of 41 for similar models. Still, it trails rivals like Claude Fable 5.2 at 65, Gemini 4 Carbon at 60, and Kimi K3.1 Pro at 46.
The firm also found it slow and wordy. It took 32.1 minutes per task, against a median of 13.3 minutes. To run the index, it wrote 127 million tokens, while the median was 78.6 million. Alibaba’s token savings were measured against its own last model, not rivals.
The model isn’t the fastest or the smartest around, but we think its prices make it worth a look for developers. You can use it now through QwenCloud’s API or its Token Plan, a subscription that works in coding tools like Cursor. Alibaba said it’s also in the QwenStudio chat site, the QwenWork app, and the Qoder coding tool.



