For K3 read this instead: https://arxiv.org/abs/2607.24653
replyThe main contribution of the K3 paper is Stable LatentMoE. Like some other models it compresses data sent between layers, which puts certain requirements on the router. K3 improves performance by using a more balanced expert selection strategy.
Compared to the Opus 5 "model card", which read like a standard Anthropic set of alignment principles and safety concerns, this presents a plethora of useful technical details that advances the state of the art.
replySame with DeepSeek papers, they are a joy to read.
replyNot an expert, but looks like they did a lot more work on the RL part (9 expert models, full sandbox access for agentic tasks, etc)?
replymost new effort in training comes in the late phase with RL techniques
replythe pretraining (slurping the internet) only goes so far, the new data being used is from human preferences and agent traces (designed and/or distilled)
Rather under-discussed back then: https://news.ycombinator.com/item?id=45766937
replyI believe OP posted it because the new Kimi K3 has 69 KDA layers (the rest are 24 Gated MLA), I think previous large Kimi models had only MLA layers.
replyIt's not the same KDA as used in Kimi Linear, though.
replyWhat's the difference? They are both called Kimi Delta Attention.
replyThe differences are explained in section 2.1.1 of the Kimi K3 technical report: https://arxiv.org/pdf/2607.24653#page=4
reply