Elon Musk has given an update on xAI's next model — and the more interesting news is not the model but the toolchain. According to Musk, pre-training of Grok 4.8 is complete and the model is now in reinforcement learning. To get there, xAI "dumped" JAX and SGLang and rewrote its entire RL and inference stack from scratch in C and C++. That rewrite, he said, "took some extra time."

For a company that spent two years on Python ecosystems and Google frameworks, that is a notable break. JAX is the de facto standard for training on TPUs; SGLang is one of the most widely used serving stacks for large models. A homegrown C/C++ stack promises more control over latency, memory and hardware utilisation — at the cost of exactly the shared libraries the rest of the industry relies on.

Timeline and tone

Musk said Grok 4.8 would be rolled out in a limited way first, including to users of the Bot offering. His own assessment was unusually candid: he said Grok 4.8 is "not as good" as Claude's Opus 5.5, calling the Anthropic model a masterpiece. A competitor publicly praising a rival's model is rare — and a sign of how tight the race at the frontier currently is.

At the same time, xAI has published no benchmark result and no confirmed parameter count. Numbers circulating on social networks cannot be verified, and rigorous comparisons are missing.

What it means for the rest of the industry

A hand-written stack is a statement: whoever controls training and inference can respond faster to hardware changes and does not have to wait on upstream projects. The price is high — custom tooling must be maintained, optimised and defended against bugs, and people who can do CUDA-adjacent work are scarce.

The real test does not arrive with an announcement but with a public model, published numbers and independent evaluation. Until then, what exists is a finished pre-training run and a promise that the foundations are finally right.