Grok 4.8 finishes pre-training as xAI rewrites its stack in C/C++, Musk says

TL;DR: Elon Musk says Grok 4.8 has finished pre-training and is now in reinforcement learning (RL), with evals "starting to climb." The team threw out JAX and SGLang and rewrote the whole RL and inference stack from scratch in C/C++, which he says cost extra time. A soft rollout to "Bot users" is planned for later this week.

What Musk said

Asked on X for an update on Grok 4.8, Musk replied early on November 10, 2026 (Elon Musk on X):

He added: "Grok 4.8 will bear no resemblance to 4.7 (thankfully). It is not yet incredible, but, this time, the foundations are built right, so I expect it to evolve to be something special."

In a follow-up reply he described the language mix as "C with a little C++, but not too much" (Elon Musk on X).

What the jargon means

• Pre-training is the big first pass where a model learns from a huge pile of text and data. Musk says that part is done.
• SFT (supervised fine-tuning) teaches the model from curated example answers. Musk says that data needed a lot of cleanup.
• RL (reinforcement learning) is the stage where the model is rewarded for better answers and gets sharper on reasoning and tasks. This is where 4.8 is now.
• JAX is Google's open-source machine learning library, and SGLang is an open-source engine for serving models. Replacing both with in-house C/C++ code means xAI now owns the software that trains and runs the model end to end.

Why it matters

Rewriting the training and serving stack in low-level code is a bet on speed and control: fewer layers between the model and the hardware, and no waiting on outside projects to fix or change things. Musk is open that it slowed this release down, and that 4.8 is "not yet incredible." The pitch is that the new foundation will let later versions improve faster.

What to watch this week

• The soft rollout. Musk said it goes to "Bot users" later this week. He did not give a date, a plan tier or a model picker name.
• Evals. He says scores are climbing in RL, but no benchmark numbers have been published yet. We will add them when xAI shares them.
• Official word. As of this post, the update comes from Musk's replies on X, not an @xai announcement, so treat timing as reported and subject to change.