Grok 4.8: C++ Rewrite Unveiled This Week

Musk Says Grok 4.8 Is Getting a Limited Rollout After xAI Rebuilds Its Stack in C/C++

Elon Musk says xAI is preparing a limited Grok 4.8 rollout after replacing JAX and SGLang with custom C/C++ code across training, reinforcement learning, and inference. He also says the rewrite took extra time, but should help deliver better performance than Grok 4.7.

What changed in Grok 4.8

The biggest update is infrastructure, not just model tuning. Musk said xAI dumped JAX and SGLang and rebuilt the RL and inference stack from scratch in C/C++, a shift meant to improve speed and control over the system.

• JAX and SGLang were replaced with in-house C/C++ code.
• The rewrite covers reinforcement learning and inference.
• xAI says this change delayed release, but it is intended to improve performance.

Why the rollout is limited

The rollout is being described as soft or limited, which suggests xAI wants to expose Grok 4.8 gradually before a broader release. Musk said the model is in reinforcement learning, that evaluation scores are improving, and that the first users to see it will likely be Grok Bot users later this week.

• Pretraining is reportedly complete.
• Reinforcement learning is the current focus.
• A small rollout is expected before wider availability.

How Grok 4.8 compares with Grok 4.7

Musk’s framing implies Grok 4.8 is meant to be a more substantial upgrade than Grok 4.7. He has said the new version is “completely different” from 4.7 and that it should be stronger, even though it is still not presented as finished or fully polished.

• Grok 4.7 is being treated as the baseline being improved on.
• 4.8 is expected to outperform 4.7 in practical use.
• xAI has not published full benchmark data for 4.8 yet.

Why the C/C++ rewrite matters

For AI infrastructure readers, the rewrite is important because custom systems can offer tighter optimization, lower overhead, and better performance tuning than general-purpose stacks. The tradeoff is engineering complexity and development time, which fits the delay Musk described.

• Custom code can improve latency and efficiency.
• Hand-built infrastructure can be tuned more aggressively than off-the-shelf tooling.
• The downside is slower deployment while the stack is rebuilt.

What users should expect next

The near-term expectation is a gradual release rather than a full public launch. If the rollout goes well, xAI could expand access after the soft launch and continue improving the model through reinforcement learning.

• Early access first.
• Ongoing evaluation during rollout.
• Possible broader availability after stability improves.

Conclusion

Grok 4.8 appears to be a major engineering reset for xAI, with Musk betting that a custom C/C++ stack will deliver better results than Grok 4.7. The model is not fully out yet, but the limited rollout signals that xAI thinks the new foundation is finally ready for users.