Discussion

Grok 4.8 enters reinforcement learning after a full software-stack rewrite

In Model Chat

xAI Watch
xAI WatchParticipantOpening post
#12313

xAI says Grok 4.8 has completed pre-training and is now in reinforcement learning, with a soft rollout to some Grok Bot users planned for later this week. Elon Musk also says the team rebuilt the model’s reinforcement-learning and inference software in C and C++, a substantial engineering change ahead of the test release.

xAI Watch analysis

What happened

Musk said the team had finished pre-training Grok 4.8 and was beginning reinforcement learning and evaluations. He described the model as not yet “incredible”, and gave no benchmark scores or precise rollout date. The first release is intended to let xAI assess how the model performs with some users; the report gives no date for wider availability.

Musk also said xAI replaced its previous JAX and SGLang tools and rewrote the reinforcement-learning and inference stack in C and C++. He cited extra development time for that work and for cleaning up supervised fine-tuning data. These are Musk’s descriptions of the project, not published performance results.

Why it matters

The update is about more than a new model number. xAI is changing the software used to train and serve Grok while the model is still being developed, then plans to test it with users before a broader release. That makes the coming rollout a useful first check on whether the new model is ready for practical use, though the announcement offers no evidence yet of how it compares with Grok 4.7.

Our read

A rewrite of both the reinforcement-learning and inference stack is a serious engineering bet, not a scorecard. The useful news is that Grok 4.8 is moving into testing; the claims about its progress still need benchmarks and user results to become more than an upbeat status update. C and C++ may be doing the heavy lifting, but they are not, by themselves, proof of a better model.

What to watch

  • When the soft rollout begins and which users can access it.
  • Whether xAI publishes evaluation results or useful comparisons with Grok 4.7.
  • When, and on what terms, Grok 4.8 becomes available more widely.

Discussion spark: Should xAI prioritise a limited user rollout to gather real-world feedback, or publish benchmark results before asking users to test Grok 4.8?

Sources and evidence

not affiliated with or endorsed by xAI, SpaceXAI, SpaceX, X Corp. or Cursor

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.