On November 10, Musk responded on social media to a user's inquiry about the release timeline for Grok 4.8. He said the team had abandoned JAX and SGLang in favor of rewriting the entire reinforcement learning (RL) and inference stack from scratch in C/C++, which explains the prolonged development period. Pretraining has been completed, and the team conducted extensive cleaning of the supervised fine-tuning (SFT) data before it basically met expectations. Currently, Grok 4.8 has entered the reinforcement learning phase, and various evaluation metrics have begun to improve. The plan is to roll it out later this week to Bot users in a small-scale gray release to observe actual performance. Musk said Grok 4.8 will be completely different from 4.7. At present, it is not yet stunning, but the right foundation has been laid, and it is expected to eventually develop into a product that stands apart from others.