Elon Musk said Nov. 10 that Grok 4.8 had entered reinforcement learning after pretraining, placing the model in a new development phase before xAI’s planned limited rollout.
Musk said the rewrite was one reason development took longer. He said xAI replaced JAX and SGLang with C and C++ systems for reinforcement learning and inference, and made extensive revisions to the supervised fine-tuning dataset.
Musk said the model’s evaluation scores are improving during reinforcement learning. He called the model a major departure from Grok 4.7 but said its current state does not impress him. EconoTimes previously covered xAI’s launches of Grok 4.7 and Grok Voice 3.0.
xAI plans to begin limited access for Bot users later this week, Musk said.
