On July 17, Kimi (Moonshot AI) revealed in its K3 technical blog that the model autonomously designed a custom chip for its own nano model. Using open-source EDA tools and a 45nm process library, K3 completed the entire chip flow – from construction and optimization to verification – in just 48 hours.

The resulting chip measures 4 mm², integrates 1.46 million standard cells, 0.277 MB SRAM, and an INT4 MAC array with fused dequantization. At 100MHz, it meets all timing requirements, delivering stable simulated decode throughput above 8,700 tokens/s – fully matching the target nano model's needs.

This concept validation highlights Kimi K3's agentic planning capability and adds a real-world case to the growing field of AI-assisted semiconductor design, potentially shortening development cycles and lowering barriers for custom silicon.

From ICgoodFind: A 48-hour chip design from a language model – that's not science fiction. The silicon works, the EDA stack is open-source, and the bar just got lower for every custom chip aspirant.