Kimi K3 jumps to 2.8 T parameters with NoPE and LatentMoE for efficiency
Kimi K3 scales the Kimi Linear model from 48 B to 2.8 T parameters, swapping RoPE for a NoPE scheme and adding LatentMoE and attention residuals for efficiency. The tweaks shave inference cost while delivering native multimodal support, marking it the largest open‑weight model to date.