Elon Musk said Grok 4.8 has entered the reinforcement learning (RL) phase after
rewriting the RL and inference stack from JAX and SGLang to C/C++. Pretraining
is complete; supervised fine-tuning (SFT) data was heavily cleaned and
evaluation metrics have begun to improve. A limited gray rollout to Bot users is
planned later this week. Musk said 4.8 will be substantially different from 4.7
and, while not yet striking, the new foundation should produce a differentiated
product over time.
