Alibaba Releases Qwen4-Max with 2.4T Parameters and Million-Token Context

Original title: Qwen4-Max首发上线千问AI平台,API服务、Token Plan同步开放
88AI Score

Alibaba unveiled its next-generation foundation model family Qwen4, launching the flagship Qwen4-Max first on the Qianwen AI platform with API and Token Plan access. Built on a new MoE architecture with Qwen Sparse Attention (QSA), the model carries 2.4T total parameters, native vision and million-token context, lifting Terminal-Bench 4.0 from 37.9 to 48.5 while cutting per-task tokens by nearly 50% and average cost by 58%. Alibaba also detailed real deployments in automotive production-line engineering and AI4S materials discovery.

Full Report

Welcome to experience it!

Today, Alibaba's new-generation foundation model Qwen4 is officially released. The Qianwen AI platform is the first to launch Qwen4-Max, with API services and Token Plan simultaneously supporting calls to the new model, fully open to developers and enterprises.

Qwen4-Max adopts a new MoE architecture based on Qwen Sparse Attention (QSA), with 2.4T total parameters, native vision, stable delivery capability for long-horizon tasks, and support for million-token context. Based on the innovative model architecture and training optimizations, Qwen4-Max achieves higher performance with lower token consumption, and continues to evolve in key capability areas such as agents and coding: in the Terminal-Bench 4.0 professional coding benchmark, at the same Xhigh thinking level, Qwen4-Max's average token usage per attempt dropped from 289K to 151K compared with the previous generation Qwen3.8-Max, a reduction of nearly 50%, while the benchmark score rose from 37.9 to 48.5, an improvement of nearly 28%. Meanwhile, thanks to a substantial drop in cache-hit pricing, benchmark results show that Qwen4-Max's average Cost per Task for completing a single task is 58% lower than that of Qwen3.8-Max.

Qwen4 is deeply optimized for real-world tasks across all domains. It can better invoke professional tools to handle long-cycle tasks and can deliver results in stages based on user feedback, advancing the deep deployment of AI in professional fields such as automotive manufacturing, game development, financial investment research, and scientific research.

In the actual production scenarios of Sanfeng Intelligent and Fanhai Tonglian, Qwen4-Max integrates drawing understanding, parametric modeling, and professional tool invocation into engineering workflows, supporting production line planning, part design, and machining verification in automotive manufacturing. At the production line level, the model uses 2D layout drawings and process parameters to invoke the existing equipment library to build a 3D production line, and carries out cycle-time simulation and PLC virtual commissioning.

At the part level, Qwen4-Max completes the modeling and design of components such as door assemblies and friction wheels, assists in machining process planning, toolpath generation, and simulation verification, and iterates machining plans based on simulation feedback, identifying issues such as over-cutting and under-cutting before machining. In the machining cases shown, the manual effort required for machining programming of a single workpiece dropped from about 475 minutes to 26 minutes, a reduction of about 95%.

For frontier exploration of human knowledge, Qwen4 has been genuinely integrated into specific AI4S fields such as materials discovery and drug discovery. In materials discovery, Qwen4 participated in the design and screening of novel coolant molecules at Deep Principle. Facing the data center cooling challenges brought by next-generation high-power chips, Qwen4 relied on the integrated dry-lab and wet-lab environment provided by Deep Principle's Mira to autonomously conduct R&D on novel coolants. During a continuous 48-hour agent run, the model completed route research, built a candidate space containing 2.7 million molecules, and sequentially carried out structural screening, MPA property prediction, first-principles calculations, retrosynthesis design, and experimental planning.

Subsequently, Qwen4 dispatched an automated wet lab to complete the preparation, synthesis, separation and purification, and testing of real materials. After 4 rounds of closed-loop iteration of "molecular design—synthesis and testing—active learning," the model ultimately obtained experimentally validated novel coolant molecules and physical samples, exploring a new materials pathway for data center cooling and demonstrating long-horizon scientific discovery capability spanning the digital and physical worlds.

In drug discovery, Qwen4-Max focused on the NLRP3 inflammasome target associated with Alzheimer's disease and autonomously carried out literature and patent retrieval, molecular design, and computational evaluation on Deep Potential Technology's PharmMaster platform, continuously optimizing based on feedback. The entire process invoked tools 454 times in total, completed 3,239 computational metric evaluations, designed and optimized 24 candidate molecules, and linked evidence retrieval, tool execution, and feedback validation into a complete scientific research chain.

The ultimately selected molecule achieved improvements over the patent starting point across multiple core computational metrics. Relative binding free energy (RBFE) calculations estimate that its nominal binding affinity to the target reached about 3.6 times that of the patent starting point. Experts at Deep Potential Technology believe that this molecule has high research value and is worth further wet-lab validation, demonstrating Qwen4-Max's ability to autonomously advance long-horizon scientific research tasks and continuously optimize candidate molecules.

More stable and longer-cycle task execution capability also enables Qwen4 to participate in optimizing the systems that support its own operation, driving AI autonomous iteration one step further. Based on Alibaba Cloud's AI infrastructure, Qwen4-Max autonomously iterates on the scheduling schemes required for its own application, inference, and training, forming a closed loop in which the model participates in optimizing its own operating system. For example, in the corresponding scenarios, the model invocation scheduling cost in data annotation decreased by an average of 9.00% compared with the reference baseline, the idle capacity duration of whole machines in GPU online allocation increased by 4.62% compared with the baseline, and the solving speed for reaching the optimal target in CPU offline rescheduling improved by more than 14.61 times. These improvements in cost and resource efficiency will also provide compute support for subsequent model iterations.

Currently, developers and enterprises can call the new Qwen4-Max model through the Qianwen AI platform's API service and Token Plan. In mainland China, the price is 10 yuan per million input tokens and 30 yuan per million output tokens, with implicit cache hits at only 0.5 yuan; on the international site, the price is 1.5 USD per million input tokens and 4.5 USD per million output tokens, with implicit cache hits at 0.075 USD. In addition, users can experience Qwen4-Max through applications such as Qianwen Office, Qoder, the Qianwen APP, and QwenStudio.

Qwen4-Max model experience address:

Qianwen AI platform:

API:

https://www.qianwenai.com/models/qwen4-max

Token Plan:

https://platform.qianwenai.com/pricing/token-plan

QwenCloud:

API:

https://www.qwencloud.com/models/qwen4-max

Token Plan:

https://www.qwencloud.com/pricing/token-plan

Qwen AI Platform - Built for Agents, Driving AI Productivity Scan the QR code below to go directly to the Qwen AI Platform experience

Click "Read Original Article" to try it out!

Read Original Article

Open in WeChat