Alibaba’s Qwen3.7-Max Redefines AI Model Efficacy with Endurance Focus

Alibaba has introduced its latest iteration of the Qwen model, named Qwen3.7-Max, which emphasizes endurance in artificial intelligence tasks rather than solely focusing on intelligence benchmarks. This shift comes as demand grows for AI models capable of maintaining performance in complex, prolonged workflows.

The Qwen3.7-Max model is now accessible through a paid API, sparking discussions within the AI community. With an Intelligence Index score of 56.6, it ranks among the highest globally, but its standout feature is the ability to execute long-running tasks effectively. Unlike previous models that excelled in initial responses, Qwen3.7-Max is designed to remain functional and useful after numerous tool interactions, making it particularly suitable for coding and office automation applications.

A key demonstration of this capability was a 35-hour autonomous kernel optimization run. Deployed on a T-Head ZW-M890 PPU, a hardware platform unfamiliar to the model during its training, Qwen3.7-Max completed 1,158 tool calls and 432 kernel evaluations, achieving a 10.0x speedup over the Triton reference. This performance underscores the model's endurance and its ability to adapt to unforeseen circumstances without prior optimization data or hardware documentation.

Endurance Over Intelligence

In AI, endurance is becoming an important metric, especially for agent systems. Traditionally, AI models have been evaluated based on their immediate intelligence and accuracy. However, the ability to sustain operations over extended periods is increasingly recognized as a key indicator of utility. Qwen3.7-Max's successful execution of the 35-hour task is a departure from models that often fail due to misinterpretations or loss of direction.

Alibaba's recent tests showed that competing models like GLM 5.1 and Kimi K2.6 fell short in similar tasks, achieving only 7.3x and 5.0x speedups, respectively. Qwen3.7-Max not only outperformed its predecessors but also outlasted them, successfully maintaining its operational loop for over 30 hours. This endurance means developers can rely less on additional structural support around the model, simplifying the integration process into various applications.

Proprietary Model with Strategic Pricing

Despite its growing reputation, Qwen3.7-Max will not be available as an open model. Alibaba has chosen a proprietary approach, providing access through its Alibaba Cloud Model Studio and Qwen Cloud. This represents a shift for developers used to the open-source nature of earlier Qwen models. The pricing structure, set at $2.50 per million input tokens and $7.50 per million output tokens, positions Qwen3.7-Max competitively against both domestic and international models.

As the AI market evolves, Alibaba's Qwen3.7-Max could set a new standard for how models are evaluated and deployed. The emphasis on enduring performance rather than immediate accuracy may promote broader acceptance of AI systems capable of managing complex tasks over extended periods. This could reshape competitive dynamics, as companies increasingly seek models that not only perform well but can also withstand the rigors of long-term application.

The launch of Qwen3.7-Max is a moment in AI development, as endurance becomes a key metric of success. With its proprietary model and strategic pricing, Alibaba is positioning itself to redefine expectations for AI models in both enterprise and startup environments.