The reinforcement learning phase for Grok 4.8 begins this week following the completion of its primary training, Elon Musk announced. The 2.5T parameter model was developed using a new C++ software stack to optimize the computational requirements and efficiency of the training process.

The transition to RL occurs once the base model's initial training cycle is complete. This update confirms the specific parameter scale and the underlying software architecture for the latest iteration of the xAI assistant.