Kimi K3 Architecture Overview and Notes | Hacker News

Hacker News users examined Kimi K3’s architecture and noted that it appears to use NoPE instead of RoPE in local layers.

Some commenters speculated that Kimi Delta’s linear-attention component may be providing the model’s positional behavior.

The discussion also questioned whether Kimi K3’s benchmark results are directly comparable to other frontier models.

Another debate focused on how much Kimi K3’s active parameter count differs from models like Opus and Fable.