Kimi K3 explained in 13min..

The newest Kimi K3, a whopping 2.8T parameter open model has been released by moonshot. Kimi K3 incorporates Kimi Delta Attention, Stable LatentMoE, and Attention Residual that all contribute to its success and hitting near Fable 5 and GPT 5.6 level in benchmarks. As MoE grows Kimi is showing just how sparse models can be scaled without sacrificing throughput and token efficiency. Follow me: X: https://x.com/calebfoundry LinkedIn: https://www.linkedin.com/in/calebeom/ TikTok: https://www.tiktok.com/@calebwritescode Chapters 00:00 Intro 01:30 Mixture of Experts 03:54 Stable Latent MoE 06:07 Kimi Delta Attention 10:23 Attention Residual 12:31 Conclusion