Kimi just dropped the model weights and technical report for K3, a 2.8 trillion parameter open-weight model with native vision and a 1M token context window. 2.5x the intelligence per unit of compute...