Alibaba launches Qwen4-Max, an AI that builds games and 3D scenes and reads whole books at once
The new model handles huge documents, runs complex workflows faster and cheaper, and helps build 3D scenes and games.

An AI program that can read up to twelve novels at once and build playable 3D games.
- Type a description of what you want created.
- The AI builds the project automatically.
- It plays-tests and refines the result.
- You get a finished usable product.
What you could do with it
- A legal or finance professional could run hundreds of pages of documents through the model at once, then complete complex workflow tasks faster and at about one-third the previous cost 1.
- A game developer could type a short description and receive a fully playable 3D adventure built in Unreal Engine, Unity, or Godot, refined through repeated automated playtesting 1.
- A 3D artist or 3D-printing designer could use Blender to model geometry, set lighting and materials, and render scenes tied together in one long context window 1.
A bigger brain with a sharper focus
Qwen4-Max is the first release in Alibaba's new Qwen4 family 2. It is not a simple all-at-once model. Instead, it uses a design called a sparse Mixture of Experts, meaning only a fraction of its total parts wake up when processing any given input 1.
Its full backbone holds 2.4 trillion parameters, but only about 71 billion fire for each token, while another 328 billion parameters help capture local word patterns at an early stage 12. That combination gives it a massive one-million-token context window and allows outputs of up to 393 thousand tokens . To picture that, a typical novel contains roughly 80 thousand tokens — so the model can hold around twelve novels in view at once .
Alibaba says pre-training went smoothly at this scale, with steady drops in error and no sudden failures during training 1. Inside tests across knowledge, coding, question answering and science subjects, the base version improved by five to thirteen percent compared with the previous Qwen3.8-Max generation while using twenty-five percent fewer active parameters .
Cost per task drops from $4.43 to $1.67
Alibaba Work360 test results
From the official announcementQwen4-Max: Your New Teammate Has Arrived
Where it wins and where it still trails
In evaluations set by Alibaba, Qwen4-Max scored first in four of sixteen benchmarks 2. Its most noticeable lead came in legal-agent work, where it reached 21.3 points on HLAB-Vals-1.1, well ahead of the next-ranked model, Kimi K3.1, at 11.7 .
Compared directly with Qwen3.8-Max, the improvement was broad 1. On the internal Work360 test covering professional tasks, the score rose from 47.74 to 52.69, and the model needed far fewer tokens to finish jobs — a twenty-four to thirty-five percent cut depending on effort level .
What people can do with it now
Alibaba is highlighting three practical abilities that the larger context window unlocks 1. First, 3D creation inside Blender, including geometry, lighting, materials and rendering for scenes used in games and 3D printing. Second, full game development across engines such as Unity, Unreal Engine and Godot, with the model able to build playable prototypes from a text prompt and refine them through repeated testing. A demo showed it building a cat-mage adventure game in Unreal Engine . Third, support for professional workflows in law, finance, healthcare, precision manufacturing, and frontier scientific research .
Costs have also moved down. Alibaba says average cost per task fell from four dollars and forty-three cents to one dollar and sixty-seven cents on Work360, a sixty-two percent drop, driven largely by cheaper pricing for inputs the model already holds in cache 12. The company has not made open weights available; the model can only be accessed through hosted services such as the QwenCloud API, the Token Plan, Qwen Studio, QwenWork and Qoder . No self-hosting or quantised versions were offered at launch .
From the official announcementBlender: 3D Modeling and Scene Creation
From the official announcementGameDev: Create Games with Real Engines
From the official announcementFDE in Action: Finance, Law & Healthcare
How good is it
According to Alibaba, Qwen4-Max ranked first on four of sixteen benchmarks and led the legal-agent benchmark HLAB-Vals-1.1 by a wide margin, scoring 21.3 against 11.7 for the next-ranked rival Kimi K3.1 2. On Alibaba's own Work360 workplace test, it rose from 47.74 under the previous generation to 52.69, and Alibaba also reported that the average cost per task fell by 62 percent, from $4.43 to $1.67 1. The company noted it beat Qwen3.8-Max by five to thirteen percent on internal knowledge, coding, question answering, and STEM evaluations while activating twenty-five percent fewer parameters . However, Alibaba admitted the new model trails behind top rivals on hard coding tasks and vision benchmarks .
score on a legal reasoning and assistance test. Higher is better. Numbers: Alibaba. Chart drawn by Herald of AI.
From the official announcementMore from the announcement
Key facts
- Qwen4-Max has a 2.4 trillion parameter backbone but only activates 71 billion per token
- It supports a one million token context window and up to 393 thousand output tokens
- Cost per task fell 62 percent, from $4.43 to $1.67 in Alibaba's Work360 tests
- It led four of sixteen Alibaba benchmarks and beat Kimi K3.1 on legal agent work
- It is available only through Alibaba hosted services, not as open weights
Sources
- Qwen
- Alibaba Releases Qwen4-Max: 2.4T MoE, 71B Active, 1M Context — marktechpost.com
If this story concerns you or your company, or uses something that belongs to you, ask for a correction or for it to be taken down. How requests work
Keep reading
Baidu's AI cloud nearly doubles, but falling ad sales pull profits down again
A fast-growing AI cloud business can't yet replace shrinking ad revenue at one of China's biggest tech companies.
Simon Willison used Claude Code to turn spare phones and tablets into a finder for a mystery beep
A developer used Claude Code to turn spare devices into a sound locator — and found an old carbon monoxide detector that urgently wanted its batteries replaced.
New modelsGoogle launches cheaper Gemini 4 Neon that matches top rivals on coding, law and scienceHerald of AI
Google launches cheaper Gemini 4 Neon that matches top rivals on coding, law and science
A new lower-cost Google AI model is competitive with pricier leaders on hard tasks, and you can start using it today.