Kimi K2 paper review

This week we will b reviewing Kimi K2 from MoonshotAI. https://arxiv.org/pdf/2507.20534 Kimi K2 is trained with agentic use in mind. During post-training, K2 undergoes a multi-stage post-training process, highlighted by a large-scale agentic data synthesis pipeline and a joint reinforcement learning (RL) stage, where the model improves its capabilities through interactions with real and synthetic environments. K2 is open weights and is available on Hugging Face. Time permitting we will also dive into the MUON optimizer used by Kimi K2.

Every week we meet (virtually) and discuss the most interesting topics in AI, ML and Deep Learning.
We usually rotate weeks so that every other week one of the members presents a deep learning/machine learning paper, frequently paired with a video explaining the concepts. In the off weeks, we present information about different projects. We cover computer vision, language (LLM's), health/hard science and generative models (Diffusion, GAN's etc.).
People of all levels of skill are welcome--from newbies to machine learning, to PhD's in DL/ML/AI.
We have been meeting for over 6 years weekly and have developed a neat pace and community--and all are welcome. Come join us and stay abreast of the biggest topics in Artificial Intelligence.