Elon Musk: Grok 4.8 is in RL and rebuilt from scratch in C/C++

Elon Musk just shared an update on xAI's next model: Grok 4.8 has officially finished pre-training, is currently deep into reinforcement learning, and its eval numbers are climbing.

A soft rollout to select Grok Bot users is slated for later this week to see how it performs in the wild.

Ditching JAX for bare-metal C/C++

The biggest pivot with 4.8 isn't just parameter scale—it is underneath the hood.

Musk revealed that xAI completely dumped JAX and SGLang, opting to rewrite their entire reinforcement learning and inference stack from scratch in C/C++. The overhaul took extra time, coupled with a massive cleanup of their supervised fine-tuning dataset to get data quality in order.

Musk added that Grok 4.8 will "bear no resemblance to 4.7 (thankfully)." While he noted it is not yet incredible, he emphasized that the technical foundations are finally built right this time around.

Why it matters

Tossing out popular AI frameworks to rebuild inference in raw C/C++ is classic Musk engineering: painful upfront, but aimed squarely at squeezing every drop of efficiency out of compute clusters.

Off-the-shelf Python-based serving stacks make rapid prototyping easy, but they often hit bottlenecks when scaling massive throughput and shaving off milliseconds. By controlling the entire stack down to bare metal, xAI looks determined to lower inference latency and unit costs.

We think this rebuild explains the recent timeline drag, but it might pay off handsomely down the road. If the foundation is as solid as Musk claims, Grok 4.8 will be worth watching closely once Bot users get their hands on it later this week.