Musk says Grok 4.8 pre-training complete, reinforcement learning under way

Elon Musk said on Tuesday that pre-training for xAI's Grok 4.8 artificial intelligence model had been completed, with the system now undergoing reinforcement learning ahead of a planned initial rollout to some users later this week.

Elon Musk said on Tuesday that pre-training for xAI’s Grok 4.8 artificial intelligence model had been completed, with the system now undergoing reinforcement learning ahead of a planned initial rollout to some users later this week.

Musk said xAI had replaced its previous software tools, JAX and SGLang, and rewritten the software stack used for reinforcement learning and inference in C and C++.

“We have dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time,” Musk wrote on his social media platform X.

https://x.com/elonmusk/status/2120088070443655190?ref_src=twsrc%5Etfw%7Ctwcamp%5Etweetembed%7Ctwterm%5E2120088070443655190%7Ctwgr%5E78bf7e1eb9a25730c292f00aeeb517d8da691eb5%7Ctwcon%5Es1_c10&ref_url=https%3A%2F%2Fwww.thenews.com.pk%2Flatest%2F1423133-elon-musk-announces-grok-48-featuring-a-full-c-stack-rewrite-what-to-know

He said the team had also needed to clean up data used for supervised fine-tuning (SFT), a stage in which an AI model is trained using curated examples.

“Now, 4.8 is in RL and the evals are starting to climb,” Musk said, referring to reinforcement learning and the model’s evaluation results. He did not provide benchmark scores.

Musk said xAI planned a “soft rollout” to Grok Bot users later this week to assess how the model performed.

He described Grok 4.8 as a significant departure from its predecessor, Grok 4.7, but cautioned that the new model was “not yet incredible”. He said he expected it to improve as development continued.

The announcement did not include a precise release date, independent performance results or details of when the model would become available to all users.