A new software framework written in C++ powered the pretraining of xAI's newest AI assistant before it entered the reinforcement learning stage. Elon Musk stated Grok 4.8 uses 2.5T parameters and is unequivocally better than the 2.1T parameter version previously trained with Jax. Data quality and training software were improved over the past few months to correct mistakes discovered during the run.

A scheduled training run for a 3T parameter model will follow using updated internal software and better data. Musk's roadmap includes Grok 4.7, which he says matches Opus 5.0 performance except for multimodal needs, and Grok 4.9, intended for the Astra and Fable class. The 2.5T parameter size of Grok 4.8 represents a 67% increase over the currently available Grok 4.6.