DeepSeek vs. Kimi: A user argues that owning compute matters

Ownership of computing resources is central to one user's case for DeepSeek over Kimi. If K2.8 is a post-trained version of K2.7—ultimately tracing back to K2 from July 2025—they see that as a possible sign that Kimi is struggling to train new base models. In a follow-up, they say work on K3 reinforcement learning is underway but argue that Moonshot is too dependent on Alibaba.