Alibaba launches Qwen4-Max for 3D design, game development and AI agents

Alibaba has released Qwen4-Max, a flagship AI model designed to run complex projects across tools such as Blender and Unreal Engine.
Built on Alibaba’s new Qwen4 architecture, the model has a 2.4 trillion-parameter backbone and a context window of 1 million tokens. It can interpret images and video alongside text, so an agent can use visual references to build a scene or inspect what’s happening in a game it’s testing. Alibaba is also targeting work that can take many steps to complete, from manufacturing plans and investment research to legal due diligence.
What Qwen4-Max can do
- 3D modeling. In Alibaba’s Blender demos, the model builds geometry, assembles scenes, applies materials and sets up lighting before rendering the result. The company also showed how those assets could be used in games or turned into physical prototypes through 3D printing.
- Game development. Alibaba demonstrated projects built in Unity, Unreal Engine and Godot. In one Unreal Engine demo, the model turned a written brief into a playable game with spells, upgrades and waves of enemies. It can run a build and use the results to refine details such as camera tracking and combat feedback.
- Video production. Paired with WonderClip, Qwen4-Max can help refine a storyboard, set up camera moves, render individual shots and assemble them into a multi-scene video.
- Agents that work around the clock. Alibaba says Qwen4-Max can power agents that plan longer assignments, keep working through revisions and deliver intermediate results for users to review. The company developed Work360 to evaluate this kind of work across professions.
Qwen4-Max vs. GPT-6.1 Sol and Claude Opus 5.5
Alibaba’s launch tests show a narrow lead on one coding benchmark and larger gaps on others.
- DeepSWE v1.1 (Agentic coding): Qwen4-Max scored 75.7%, just ahead of Kimi K3.1 at 75.5%, GPT-6.1 Sol at 75.2% and Claude Opus 5.5 at 74.2%.
- Agents’ Last Exam v1.1 (Long-horizon valuable tasks): Qwen4-Max scored 46.7% on longer tasks, up from 27.9% for Qwen3.8-Max and ahead of Claude Opus 5.5 at 43.4%.
- Terminal-Bench 4.0 (Agentic terminal coding): Qwen4-Max scored 48.5% on tasks carried out through a terminal, behind Claude Opus 5.5 at 59.6% and Kimi K3.1 at 67.2%.
In independent testing, Vals AI recorded 21.25% for Qwen4-Max on its legal-work benchmark, compared with 9.17% for Claude Opus 5.5. On Artificial Analysis’s broader Intelligence Index, Qwen4-Max scored 43, behind GPT-6.1 Sol at 55 and Claude Opus 5.5 at 59 when both rivals ran at their maximum reasoning settings.
Qwen4-Max pricing and availability
QwenCloud charges $1.50 per million input tokens and $4.50 per million output tokens. At standard API rates, generating the same number of output tokens costs less than half as much as with GPT-6.1 Sol and less than a quarter as much as with Claude Opus 5.5.
- GPT-6.1 Sol: OpenAI charges $2 per million input tokens and $10 per million output tokens for prompts up to 272,000 tokens.
- Claude Opus 5.5: Anthropic charges $4 per million input tokens and $20 per million output tokens.
Qwen4-Max also used fewer tokens than its predecessor in Alibaba’s tests. It used up to 49% fewer tokens than Qwen3.8-Max on Terminal-Bench 4.0. In a separate Work360 evaluation, Alibaba reported that lower token use and lower API rates brought the average cost per task down 62%, from $4.43 to $1.67.
Qwen4-Max is available in the Qwen app. Developers can access it through QwenCloud’s pay-as-you-go API under qwen4-max or through a Token Plan subscription.