Open models recap: more on Kimi K3, Qwen 3.8, Xi's WAIC speech, distillation, the open-closed gap, and what's next

In the latest quarterly open model roundup, Nathan Lambert and Florian Brand discuss the recent releases of AI models and the current state of the open model ecosystem. They begin by talking about the release of Kimi K3, which seems to be accelerating the pace of development in open models. The conversation then moves on to the GLM 5.2 model, which continues to play a significant role in the open model landscape.

The pair then delve into the performance of Chinese models in comparison to their US counterparts, analyzing the geopolitical implications of this competition.

The discussion also covers the data and environments used to train these models, as well as a tour of the Chinese labs involved in their development. They explore the differences between the Chinese providers, such as Qwen, DeepSeek, and MiniMax. In addition, they touch upon the US open-model ecosystem and the cybersecurity concerns surrounding open models, particularly in the context of the recent speech by Chinese leader Xi on the importance of openness and open-source strategies.

A major topic of conversation is the concept of distillation and its potential to improve the performance of open models. Nathan and Florian discuss the challenges and limitations of this technique, particularly in the case of Kimi K3, which requires significant computational resources to load the weights and fine-tune the model. They also share their experiences using Kimi K3 and provide insights into its capabilities and limitations.

Finally, the pair share their predictions for the future of open models and the frontier tier list, offering their thoughts on what to expect in the coming months. The conversation concludes with a call to action for listeners to order Nathan's book on post-training knowledge and to explore the course on educational post-training videos he is putting together.