Musk: Grok 4.8 completes pretraining, small-scale rollout to begin later this week

November 10 — Musk posted an update on Grok 4.8's progress: pretraining is complete, it is now in the reinforcement learning (RL) phase, and evaluation scores are starting to rise. xAI plans a soft rollout to Bot users later this week first. Musk said the team has abandoned JAX/SGLang and rewritten the entire reinforcement learning and inference system from scratch in C/C++, which is one reason for the delay; supervised fine-tuning (SFT) data also underwent extensive cleanup. He called Grok 4.8 "completely different" from 4.7 and said it is "not yet great," but the underlying foundation has been laid.