Share Your Story
Share Your Story
Artificial Intelligence November 16, 2026

Alibaba Launches Qwen4-Max, a New Flagship AI Model Built to Cut the Cost of Agent Work

Alibaba has launched Qwen4-Max, its flagship AI model with a 2.4 trillion parameter backbone, a one million token context and a 62% lower cost per task.

Alibaba has introduced Qwen4-Max, the new flagship model in its Qwen family, pitching it as an AI teammate for work, creativity and discovery. The launch, announced on 16 November 2026, follows a short preview that opened on 11 November and puts the model on general release through the QwenCloud API.

Cheaper to run, not just bigger

The headline claim is about cost, not size. On the Qwen team’s own Work360 evaluation, the average cost per task fell from $4.43 with Qwen3.8-Max to $1.67 with Qwen4-Max, a reduction of 62%. Token use fell by about 24% at the highest effort setting and 35% at high effort.

The savings show up on coding work too. On Terminal-Bench 4.0, token use fell by about 48% and 49% at the two effort settings, while the model’s Work360 score rose from 47.74 to 52.69. Work360 spans 721 professional fields across 26 domains, from finance to healthcare.

API pricing matches the cheaper running costs. Qwen4-Max costs $1.50 per million input tokens and $4.50 per million output tokens on QwenCloud, against $2 and $6 for Qwen3.8-Max. Implicit cached input costs $0.075 per million tokens, down 70% from the previous generation.

What the model is built on

Qwen4-Max runs on a new Qwen4 architecture. It is a sparse mixture of experts model with a 2.4 trillion parameter backbone, 71 billion parameters activated per token and a further 328 billion N-gram Embedding parameters. It reads text, images and video, and holds a one million token context window.

Training at that scale stayed stable, according to the Qwen team. The loss curve fell smoothly with no spikes through the run, and the base model beat Qwen3.8-Max on in-house tests across knowledge, code, question answering and science while activating 25% fewer parameters.

The demos go beyond chat

Alibaba framed the launch around long-running work rather than chat. The model builds 3D scenes in Blender and playable games in Unity, Unreal Engine and Godot from a single brief, then plays them back and refines them from visual feedback.

In factories, the examples are concrete. At SANF and Pacific Union, the model planned production lines and designed machined parts, and the manual effort to program a single workpiece fell from about 475 minutes to 24 minutes.

The professional demos stretch further. In finance it pulls market data through 90 tools and returns workbooks and slide decks. In legal work it sorted 198 files into 14 folders and produced an 86-page due diligence report, and in healthcare it reviewed 1,433 medical images across a six-day case.

The science demos reach furthest. In a continuous 48-hour run on Deep Principle’s Mira platform, the model screened 2.7 million molecules and directed an automated wet lab through four design and testing loops to produce a physical sample of a new coolant material for data centre chips.

A drug discovery demo targeted the NLRP3 inflammasome, a driver of neuroinflammation linked to Alzheimer’s disease. The model made 454 tool calls and 3,239 evaluations across 24 candidate molecules, and its candidate showed about 3.6 times the estimated binding affinity of the reference molecule. Wet lab validation is still pending.

What independent testing shows so far

Outside Alibaba’s own numbers, the first independent score is encouraging but modest. Artificial Analysis puts Qwen4-Max at 43 on its Intelligence Index, above the median of 41 for comparable models, and confirms the $1.50 and $4.50 per million token prices.

The same independent run flags a trade-off. Qwen4-Max generated 127 million output tokens across the index, well above the 78.6 million median, so it writes more to reach its answers.

Alibaba’s own comparison chart shows it scoring 21.3 on the HLAB legal test, ahead of Anthropic’s Opus 5.5 at 9.2, while trailing on Terminal-Bench 4.0 with 48.5 against 59.6 for Opus 5.5.

That mix is the honest picture. The cost and efficiency claims come from Alibaba’s own evaluation, the demos are single showcases run with partner platforms, and the drug candidate still awaits wet lab validation. Max itself is closed, though Alibaba says smaller Qwen4 models will be released as open weights.

How to try it

Qwen4-Max is live now on QwenCloud, through the pay-as-you-go API under the model id qwen4-max and through the Token Plan subscription. It is also available in Qwen Studio, QwenWork and Qoder, Alibaba’s own apps and coding tools.

The general release closes a short preview window. Qwen4-Max first appeared on 11 November as a preview available only through Token Plan, with no per-token price, and that preview model is retired after 25 November. The new release adds the API and published prices.

“Your new teammate has arrived. Ready when you are.”

Qwen Team, Alibaba

The pitch here is not a smarter chatbot. It is cheaper, steadier agent work, with the model doing more of the job for less money and fewer tokens. Whether that holds outside Alibaba’s own benchmark is the question every team running agents at scale will now test for itself.