Qwen4 Max is here. Alibaba’s new AI teammate brings 2.4 trillion parameters, native vision, and smarter long-task skills for work and discovery.
Alibaba’s Qwen team has launched Qwen4 Max, calling it a powerful new AI teammate for everyday work, creative projects, and serious discovery. Released in mid-November 2026, Qwen4 Max brings a fresh architecture, native vision skills, and a massive one-million-token context window. It aims to help people finish complex tasks faster and with better results.
Key Features Highlights
- 2.4T parameter backbone
- 71B active parameters
- Native vision support
- 1 million token context
- Stronger tool use
- Lower token costs
- Live Delivery agents
- Real manufacturing help
- Science discovery power
- Closed weights, cloud access
What Makes Qwen4 Max Different?
Qwen4 Max stands out because it handles long projects that used to need constant human oversight. It works across coding, design, manufacturing, finance, law, healthcare, and even scientific research. The model keeps the big picture in view and delivers usable results step by step.
Core Specs of Qwen4 Max
Qwen4 Max uses a sparse mixture-of-experts design. It has a 2.4-trillion-parameter backbone. Only 71 billion parameters activate for each token. An extra 328 billion N-gram embedding parameters add local pattern knowledge. The model accepts text, images, and video. Its context window reaches one million tokens.
New Architecture Inside Qwen4 Max
The Qwen4 series introduces clear upgrades. Gated DeltaNet layers compress past information into fixed-size states. Qwen Sparse Attention then picks the most relevant parts at a fine grain. Gated Residual connections use four parallel paths with dynamic gates. These changes improve efficiency, capacity, and training stability. Pre-training ran smoothly with no loss spikes.
Efficiency Gains and Lower Costs
Qwen4 Max does more with fewer tokens than its predecessor. On internal work tests, token use dropped by about 24 to 35 percent. On coding-style benchmarks the drop reached nearly 50 percent. Average cost per task fell from $4.43 to $1.67 — a 62 percent cut. API prices also dropped, especially for cached input.
Strong Performance on Real Work
The team built a broad evaluation called Work360 that covers 26 major domains and hundreds of specialized fields. Qwen4 Max scored 52.69, up from 47.74 on the previous top model. It shows better coverage of industry tasks and stronger collaboration skills.
3D Modeling Strengths
Qwen4 Max works smoothly inside Blender. It handles geometry, scene layout, materials, lighting, and rendering. The model creates architecture, interiors, landscapes, and science-fiction scenes. Finished assets can move into games or 3D printing.
Game Development with Real Engines
The model supports Unity, Unreal Engine, and Godot. From a simple text brief it builds scenes, gameplay logic, materials, HUD elements, and effects. It tests the game, watches visual feedback, and improves camera work, animation, and pacing. Examples include a cat-mage adventure, a flying fantasy title, and a mobile racer set around West Lake.
Help for Automotive Manufacturing
In real factory settings Qwen4 Max plans production lines, designs parts, and prepares machining paths. It builds 3D models from 2D layouts, simulates cycle times, and checks for problems before cutting metal. In one case, programming time for a single workpiece dropped from about 475 minutes to 24 minutes.
Agentic Video Production
Paired with tools like WonderClip, Qwen4 Max can manage full video pipelines. It refines scripts, creates consistent assets, plans camera moves, renders clips, and edits multi-scene films. Natural language plus controlled previsualization keeps results predictable.
Support Across Finance, Law, and Healthcare
In finance the model pulls market data through dozens of tools, runs multi-factor analysis, and delivers Excel files plus PowerPoint decks with clear source links. In legal due diligence it organizes hundreds of messy documents, cross-checks public records, and produces long, fully cited reports. In critical care it reviews large medical records, images, and notes to support clinical decisions.
Long-Running Agents That Stay Under Control
Qwen4 Max powers agents that work for hours or days. A feature called Live Delivery shares early usable versions so users can guide the work in real time. An open-source plugin helps teams adopt this style. The model has also optimized cloud resource scheduling inside Alibaba Cloud, cutting costs and improving GPU and CPU use.
Advances in Scientific Discovery
In one 48-hour run the model designed, screened millions of molecules, planned experiments, and directed a wet lab to produce a new coolant material for high-power chips. In drug research it searched literature, designed candidates, and prioritized a molecule estimated to bind more strongly to a key target linked to Alzheimer’s-related inflammation.
Safety and Responsible Design
Safety checks run throughout training and use. Secure coding performance improved. Cybersecurity training focuses on defense rather than attack skills. Special data and post-training reduce risks around sensitive topics. The team stresses human oversight for high-stakes decisions and lawful data use.
How to Start Using Qwen4 Max?
Qwen4 Max is available now through QwenCloud via API or Token Plan. Users can also try it inside Qwen Studio, QwenWork, and Qoder. Weights for the Max model remain closed. Smaller models in the Qwen4 family are expected to arrive later with open weights.
What This Means for Users Everywhere?
Whether you work in design, engineering, research, or office tasks, Qwen4 Max offers a practical partner that stays on the job. It turns complex goals into concrete deliverables while keeping people in charge. The lower cost and higher token efficiency make advanced capabilities more accessible across countries and team sizes.
Qwen4 Max marks the start of the Qwen4 series. It combines strong reasoning, vision, and long-horizon planning into one teammate ready for real projects. As more models in the family appear, the same techniques will reach even more users.



