Alibaba has unveiled Qwen4-Max, a flagship AI model with built-in image and video understanding, a one-million-token context window, and tools for projects that can run for hours. Alibaba showed it building a playable game, planning factory production lines, and carrying out research. In the company's Work360 benchmark, the average cost per task fell from $4.43 to $1.67 compared with Qwen3.8-Max.

This release marks a significant expansion in enterprise artificial intelligence capabilities, offering high-level reasoning and cost efficiency across multiple industries. Businesses and developers are increasingly looking toward large-scale models that can execute complex, multi-stage workflows without constant human intervention.

How Qwen4-Max Achieves Higher Scores and Fewer Tokens

Qwen4-Max scored 52.69, up from 47.74, on Work360, which covers 26 broad domains, 182 subdomains, and 721 specialized fields. At Xhigh and High reasoning effort, it used 24% and 35% fewer tokens on Work360 and 48% and 49% fewer on Terminal-Bench 4.0, respectively.

Its 2.4-trillion-parameter backbone activates 71 billion parameters per token, with another 328 billion dedicated to recognizing nearby token patterns. Gated DeltaNet and sparse attention help it process long inputs, while four gated residual branches control how information moves between layers. Alibaba reported stable training and 5% to 13% lower bits per byte across its knowledge, code, question-answering, and STEM tests, with 25% fewer active parameters than the previous base model. Two competing checkpoints underwent additional training, which may affect the comparison.

Understanding the architectural foundation helps clarify why large language models consume fewer computational resources while delivering higher accuracy. By activating only a fraction of its massive parameter count per token, the system balances raw power with operational efficiency.

Capabilities in Games, Video, and Factory Automation

Alibaba showed the model building and rendering 3D scenes in Blender for games and 3D printing. It supports Unity, Unreal Engine, and Godot, and built Meow of Magic, a playable Unreal Engine game. It can use feedback from test runs to refine gameplay and visuals. With WonderClip, it took a video from storyboard and camera planning through rendering and editing.

At SANF and Pacific Union, the model helped turn 2D layouts into 3D automotive production lines, simulate cycle times, test controllers, design parts, and check machining paths. In one demonstration, manual programming time for a part fell from 475 minutes to 24 minutes.

These practical applications demonstrate how digital design tools integrate with advanced reasoning systems to automate repetitive engineering tasks. Manufacturers can test virtual factory setups before committing physical resources, reducing downtime and engineering overhead.

Applications in Finance, Law, and Healthcare

  • Finance: Qwen4 used 90 tools to gather market data, cross-check disclosures, and analyze CSI 300 derivatives. It produced an Excel workbook with source links and an editable PowerPoint deck tied to the workbook.
  • Law: It sorted 198 pre-IPO working papers into 14 folders, checked public records, and produced an 86-page report covering 12 review areas and 14 legal issues, with citations to the underlying documents.
  • Healthcare: In a critical-care demo, it reviewed six days of records, including 1,433 medical images. Alibaba said its imaging finding matched the attending doctor's diagnosis and its four suggested actions aligned with the doctor's orders.

Professional services often require processing vast amounts of unstructured data under tight deadlines. Automated agents capable of handling multi-step analysis across legal, medical, and financial domains change how firms manage large documentation projects.

How Multi-Stage Agents Work in Practice

The open-source Live Delivery plugin lets agents hand over an early version of their work, then revise it as users give feedback. Alibaba developed the model alongside its agent software and products such as Qoder. In cloud scheduling tests with the MindOpt solver, Alibaba reported 10.42% lower model-call costs for data annotation, 4.62% more time with entire GPU machines available, and a more than 14.61-fold speedup in solving an offline CPU scheduling problem.

Task delegation and iterative feedback loops allow artificial intelligence systems to collaborate effectively with human workers. Instead of delivering a single final output, stage-based agents present intermediary results that can be adjusted continuously.

From Simulation to Scientific Laboratory Testing

In a 48-hour run on Deep Principle's Mira platform, Qwen4 screened 2.7 million molecules and produced a coolant sample for high-power chips after four rounds of design and testing. In a separate Alzheimer's research demo on PharmMaster, it made 454 tool calls, ran 3,239 computational evaluations, and developed 24 candidates. Simulations put the selected molecule's estimated binding affinity at 3.6 times that of the reference compound. It still requires lab validation.

Computational biology and materials science benefit greatly from high-throughput screening capabilities. By automating molecular discovery and evaluation cycles, researchers can explore millions of chemical combinations in a fraction of traditional timelines.

Safety Standards and Pricing Structure

In Alibaba's secure-coding tests, the share of functions that passed both functionality and security checks rose from 22% to 31.8%. The share that passed security checks across functions, files, and repositories rose from 8% to 14.7%. Alibaba also reported stronger defensive cybersecurity performance and better handling of risky chemical, biological, radiological, and nuclear requests in internal evaluations.

Qwen4-Max is available through the QwenCloud API, Token Plan, Qwen Studio, QwenWork, and Qoder. The API supports web search, tool calling, and structured outputs. It costs $1.50 per million input tokens, $4.50 per million output tokens, and $0.075 per million cached input tokens, with the implicit cache rate down 70% from Qwen3.8-Max. Alibaba plans to release the weights of smaller Qwen4 models later.

Maintaining rigorous safety standards while expanding commercial availability remains a primary focus for AI developers. Transparent pricing tiers and flexible deployment options ensure that organizations of various sizes can access advanced reasoning infrastructure.