I think one crucial part that everyone seems to have missed is that they replace the previous Gated MLA layers (Kimi K3) with DSA (K3.1) via post-training to reduce the inference cost. In theory it should lower the kv-cache pressure by a huge magnitude scale.