Alibaba has launched Qwen4-Max, a flagship AI model aimed at professional work in industry, science, and finance. The mixture-of-experts model has 2.4 trillion parameters but activates 71 billion per token, according to the Qwen team. It processes images natively and has a one-million-token context window.
Qwen4-Max introduces the revised Qwen4 architecture. Three layers compress earlier context into a condensed state. A fourth selects the relevant parts, reducing the computation needed for long inputs. The design also uses four parallel connections between layers with dynamic gates. An N-gram embedding layer with 328 billion parameters provides common token sequences as building blocks. Alibaba says training remained stable throughout the run.
In Alibaba's internal evaluations, Qwen4-Max-Base scored best among five checkpoints tested across knowledge, code, question answering, and STEM. It improved 5 to 13 percent over Qwen3.8-Max-Base while activating 25 percent fewer parameters per token. On Alibaba's Work360 benchmark, which covers tasks in 721 specialized fields, Qwen4-Max used roughly a quarter to a third fewer tokens than its predecessor. Alibaba says the average cost per task fell 62 percent, from $4.43 to $1.67, while the score rose from 47.74 to 52.69.
Read nextVercel adds Alibaba’s Qwen4-Max to AI GatewayBenchmark results vary by task
External results are mixed. Qwen4-Max scored 46.7 points on Agents' Last Exam v1.1, the highest result in its comparison group. It trailed Opus 5.5 and the GPT models on Humanity's Last Exam. On HealthBench Professional, it finished behind the GPT models but ahead of Opus 5.5.
At its highest reasoning setting, Qwen4-Max scored 43 on the current Artificial Analysis Intelligence Index, nine points above Qwen3.8-Max and three below Kimi K3.1 Pro. Artificial Analysis put its cost at $3.27 per index task, compared with $10.02 for Qwen3.8-Max and $3.68 for Kimi K3.1 Pro. Qwen4-Max used about 105,000 output tokens per task, down from 171,000 for its predecessor. Through Alibaba's API, it costs $1.50 per million input tokens and $4.50 per million output tokens. Cache hits cost $0.075.
On Terminal-Bench 4.0, Qwen4-Max scored 48 percent, behind Kimi K3.1 Pro's 67 percent. It tied Kimi at 26 percent on Cognition's FrontierCode 2.0 coding benchmark and trailed Tencent Hy4 at 31 percent. Hy4 costs $1.03 per Artificial Analysis index task but scores eight points below Qwen4-Max.

Alibaba shows the model at work
Alibaba designed a feature called "Live Delivery" for long agent tasks. It produces a usable first version early, then improves it in later passes while users can change course. Alibaba has open-sourced the accompanying Qwen-Live-Delivery-Plugin.
Alibaba says automotive partners SANF and Pacific Union cut manual programming time for one workpiece from about 475 minutes to 24 using Qwen4-Max. In finance, the model uses 90 MCP tools to gather market data and produce Excel and PowerPoint files with numbers traceable to their sources. Alibaba also says the model and its MindOpt solver reduced the cost of allocating compute jobs for data annotation by an average of 10.42 percent.
In a lab demonstration with Deep Principle's Mira system, Qwen4-Max ran autonomously for 48 hours, according to Alibaba. It searched 2.7 million molecular candidates and completed four cycles of designing, synthesizing, and measuring a coolant for data center chips. In work with DP Technology's PharmMaster platform, Alibaba says a candidate compound bound to a target protein about 3.6 times better than a patented starting compound. DP Technology's experts plan further lab tests.
Alibaba reports that functional and secure code rose from 22 to 31.8 percent on Last Secure Code Bench. Its internal CBRN review found no capacity for innovation that would significantly increase a user's capability, the company says. Qwen4-Max is available through QwenCloud, QwenWork, Qoder, and QwenStudio. Alibaba plans to release smaller Qwen4 models as open source later.



