Alibaba’s Qwen4-Max is designed to handle demanding professional tasks while using fewer resources than its predecessor. The company highlights lower benchmark costs, long-context processing and agents that let users steer work before it is finished.
A design aimed at long inputs
Qwen4-Max is a mixture-of-experts model with 2.4 trillion parameters, of which 71 billion are activated for each token, according to the Qwen team. It processes images natively and supports a one-million-token context window.
Its Qwen4-series architecture combines context compression with selective retrieval. Three layers condense earlier context into a shorter state, and a fourth selects the relevant parts. This approach reduces the work required to process long inputs, since the model does not treat every part of the context equally.
The design also includes four parallel connections between model layers, dynamic gates intended to stabilize information flow, and an N-gram embedding layer with 328 billion parameters. Alibaba says training the 2.4-trillion-parameter configuration was stable, with training loss declining without sudden spikes.
In Alibaba’s internal evaluations, Qwen4-Max-Base performed best among five tested checkpoints. The company reports a 5 to 13 percent improvement over Qwen3.8-Max-Base while activating 25 percent fewer parameters per token.
Lower costs, with uneven benchmark results
Alibaba’s Work360 benchmark covers industrial tasks across 721 specialized fields. The company says Qwen4-Max used roughly a quarter to a third fewer tokens than its predecessor on the same tasks. On Terminal-Bench 4.0, token use fell by nearly half.
Alibaba reports that the average Work360 task cost dropped 62 percent, from $4.43 to $1.67, while the benchmark score rose from 47.74 to 52.69. These are company-reported results, and performance elsewhere varies by task.
On Agents' Last Exam v1.1, a benchmark for long-running agent tasks, Qwen4-Max scored 46.7 points, the highest result in the comparison group. It scored below Opus 5.5 and the GPT models on Humanity's Last Exam, while on HealthBench Professional it was behind the GPT models but ahead of Opus 5.5.
At its highest reasoning setting, the model scored 43 points on the Artificial Analysis Intelligence Index, nine more than Qwen3.8-Max and three behind Kimi K3.1 Pro. Artificial Analysis puts Qwen4-Max at $3.27 per task on that index, compared with $10.02 for Qwen3.8-Max and $3.68 for Kimi K3.1 Pro. The cost advantage reflects lower token prices, cheaper cache hits and fewer output tokens per benchmark task.
Results on other comparisons are less favorable. Qwen4-Max scored 48 percent on Terminal-Bench 4.0, below Kimi K3.1 Pro at 67 percent. On FrontierCode 2.0, it tied Kimi at 26 percent and trailed Tencent Hy4 at 31 percent. Alibaba’s API prices are $1.50 per million input tokens, $4.50 per million output tokens and $0.075 for cache hits.
Agents that show work before it is done
Alibaba’s “Live Delivery” feature is intended to make autonomous work easier to supervise. Rather than waiting for a long task to finish, users get an early usable version that the model expands in later passes. They can review progress and change direction while the task is running.
The company says automotive partners SANF and Pacific Union used Qwen4-Max to reduce manual programming time for a single workpiece from about 475 minutes to 24. In finance, the model uses 90 MCP tools to gather market data and produce Excel and PowerPoint files with figures traceable to their sources.
Alibaba also describes internal uses in cloud operations. Working with the company’s MindOpt solver, Qwen4-Max cut the cost of allocating compute jobs for data annotation by an average of 10.42 percent, according to Alibaba. For distributing CPU tasks in data centers, it found optimal solutions more than 14 times faster than the comparison method.
From lab experiments to game development
Alibaba says Qwen4-series models have been paired with Mira, a system from Deep Principle, to run a lab experiment autonomously for 48 hours. The project searched 2.7 million molecular candidates and completed four cycles of designing, synthesizing and measuring a coolant for high-performance data center chips.
In another research example, Qwen4-Max worked with PharmMaster from DP Technology on a compound related to an inflammatory response linked to Alzheimer's disease. The model helped optimize 24 candidate molecules. Alibaba says the final candidate bound to the target protein about 3.6 times better than the known starting compound from a patent; DP Technology experts considered it promising enough for further lab tests.
The company also presents workflows for Unity, Unreal Engine and Godot that start from a natural-language game description and extend through scenes, logic, assets, interface and effects. During development, the model can run the game and use visual and interactive feedback to adjust elements such as camera work, animations and lighting. Alibaba’s example is “Meow of Magic,” an Unreal Engine adventure featuring a cat wizard. The company says Qwen4-Max can also create Blender 3D models for reuse in games or 3D printing.
Together, the examples position Qwen4-Max as a lower-cost model for sustained work, with user oversight built into some agent workflows. Its benchmark standing depends on the task, and the application claims described here come from Alibaba and its partners.