Alibaba AI Optimizes Chip Code Autonomously: 10x Speedup

Alibaba's new AI model, Qwen3.7-Max, optimized code for its own custom chip autonomously for 35 hours. This model, released by the Qwen team, is designed for complex software projects and independent work over long periods. What's interesting is that Qwen3.7-Max performed a kernel optimization for a hardware-based attention kernel. It achieved an average 10x speedup over the original code. The model started with no prior knowledge of the chip architecture or hardware documentation. During this 35-hour test, the model ran 432 kernel tests and made 1,158 tool calls. It compiled, measured, and revised the code, catching errors and finding performance bottlenecks on its own. Other models showed lower speedups in the same setup. The bottom line is this model can significantly enhance efficiency in hardware optimization, potentially speeding up development for specialized computing.