Alibaba's latest AI model ran autonomously for 35 hours to optimize code for its own custom chip

Alibaba’s Qwen team launched Qwen3.7-Max exclusively via Alibaba Cloud Model Studio.

In a benchmark, the model autonomously debugged and optimized a hardware attention kernel for 35 hours on T-Head-ZW-M890 accelerators.

It made 432 tests and 1,158 tool calls, with a reported average 10x speedup over the reference code.

The result suggests Qwen3.7-Max can handle long-running agentic coding and systems optimization tasks.