Open models recap: more on Kimi K3, Qwen, distillation, and what's next

This piece is a quarterly discussion between AI researchers Nathan Lambert and Florian Brand focusing on the current state and rapid acceleration of open-weight artificial intelligence models. The conversation centers heavily on the recent release of the Chinese model Kimi K3, alongside developments from other providers like Qwen, DeepSeek, GLM, and MiniMax, while exploring the broader economics, geopolitics, and technical realities shaping the industry.

Key topics include the ongoing debate over how far open models lag behind closed frontier systems, with an emphasis on performance in agentic coding and complex tasks. The participants examine why Chinese labs are producing such competitive models efficiently, citing factors such as focused research teams, capital efficiency, rising domestic chip production, and improving access to data environments. They also discuss the practical challenges of deploying and fine-tuning massive new open models due to heavy infrastructure demands and swamped APIs.

This discussion matters because the gap between closed-source industry giants and open-weight models is not widening as many expected, despite heavy American export restrictions. Understanding how global labs iterate rapidly, share open weights, and optimize post-training techniques provides a clear picture of where the AI landscape is heading and how developers actually utilize these powerful tools in practice.