Elon Musk says Grok 4.8 will finish pretraining this week and enter reinforcement learning

Musk said Grok 4.8 will finish its initial training in the week of Sept. 14 and start RL; reinforcement learning and subsequent safety evaluation will determine when the model reaches users.

In this brief: 2 sections 1 min read

- Elon Musk posted on X saying Grok 4.8 is a 2.5T model and will finish pretraining the week of Sept. 14.
- Musk said the run used a new internal C++ training stack rather than the previously used JAX-based tooling.
- He committed to starting reinforcement learning (RL) during that week; no release date or API details were provided.
- Reinforcement learning can materially change model behavior for reasoning, coding, and agentic tasks.
- The RL duration, safety checks, and deployment plans will determine when xAI makes the model available to users.
- Musk's claim is public progress but not a shipped model or an API endpoint.

Read full analysis on runtimewire.com

Useful?