The latest iteration of Elon Musk’s chatbot is currently in the reinforcement learning phase after xAI overhauled its core technical architecture. The company discarded the JAX and SGLang frameworks and rebuilt its inference and reinforcement learning (RL) stacks in C/C++ to provide a more stable base for the model. Pre-training for the newest version is finished, and the developer has completed a cleanup of the supervised fine-tuning data.
The Grok 4.8 model will be available to Bot users through a soft rollout starting later this week. Musk said the new model bears no resemblance to the previous 4.7 version and, while noting the model is not yet “incredible,” stated that its evaluation scores are beginning to climb.
