“The K2.5 model was trained on a cluster of NVIDIA H800 GPUs where each node has 2 TB of RAM and 8 GPUs connected by NVLink.”