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

Editor's take

Alibaba's Qwen team has demonstrated a significant advancement in autonomous AI by having its Qwen3.7-Max model operate independently for 35 hours, successfully optimizing code for Alibaba's proprietary AI chips. This achievement moves beyond typical LLM inference, showcasing a model capable of sustained, goal-directed operation in a complex, real-world development environment.

This development matters because it signals a tangible step towards more sophisticated AI agents that can perform intricate tasks without continuous human oversight, potentially accelerating hardware-software co-design and improving efficiency for specialized AI silicon. Its performance, matching Claude Opus 4.6 and surpassing domestic competitors like DeepSeek V4 Pro and Kimi K2.6, positions Alibaba as a strong contender in the global LLM race, particularly for enterprise applications.

Future developments to monitor include the scalability and robustness of these autonomous agents across diverse coding tasks and hardware architectures, as well as the security implications of AI models operating with such extended autonomy. Understanding the specific optimizations achieved and their impact on chip performance will be crucial in assessing the true value of this breakthrough.