Grok 4.8 Is Coming This Week After a Full Rebuild, Says Musk

Grok 4.8 has finished its first stage of training and is now in reinforcement learning, Elon Musk said, after his team at SpaceXSI rewrote the software that trains the model and generates its answers from scratch in C and C++.

Reinforcement learning is a later training step that rewards the model for better answers. Musk said the team dumped JAX and SGLang, common tools for training and running large AI models, and that the rewrite took extra time. A lot of cleanup was also needed on the example data used to teach the model how to respond, work he said was required “to get that mostly right.”

“Now, 4.8 is in RL and the evals are starting to climb,” he wrote on X early Tuesday. He said a limited rollout to Bot users is coming later this week, “and see how it goes.”

He later described the new code as “C with a little C++, but not too much,” and said there is even some assembly, a low-level form of code, in the mix.

Musk said the new software used to train Grok and generate its answers is based on work by the engineers who built Starlink. At SpaceX, “we write our own headers and calculate our own time (includes leap seconds).”

In other words, the company builds some of the most basic parts of that software itself, including how the system keeps time.

“One day, super intelligence will write a better software stack, but that day has not yet come,” Musk wrote. “This is tight code.”

Musk said Grok 4.8 will be a sharp break from today’s model. “Grok 4.8 will bear no resemblance to 4.7 (thankfully).”

The new version is “not yet incredible,” he added, but “the foundations are built right,” and he expects it to “evolve to be something special.”

That update follows the roadmap Musk shared in mid-September, when he said Grok 4.8 would finish training that week and start reinforcement learning. He also described 4.8 then as a 2.5 trillion-parameter model trained with a new C++ software stack.

In a separate reply Tuesday, asked how Grok 4.8 would compare with Anthropic’s Opus 5.5, Musk said it was “not as good” and called Opus 5.5 a “masterpiece.”

He added that Grok 5 “should exceed Opus 5.5 and maybe that happens in 2 or 3 months, with continued acceleration thereafter.”

He did not say when a wider public release might come. Grok 4.8 is still in reinforcement learning, and Musk said the model’s benchmark scores are starting to improve.