Elon Musk: Grok 4.8 Is Several Model Variants in Beta, With No Pick Yet for Wide Release
@firesidealpha “Grok 4.8” is currently several different model variants in beta. We’re not yet sure which of the variants will go to wide release.
Context
It confirms xAI's next Grok model is already being tested on users in several versions, so people using Grok Bot may be talking to unreleased models without knowing which one.
Elon Musk replied to a Fireside Alpha clip that Grok 4.8 is currently several different model variants in beta, and that xAI is not yet sure which one will go to wide release. In the clip, from his DealBook interview, Musk said Grok 4.8 is in the reinforcement learning phase and that Grok Bot users sometimes get different versions of Grok being tested. In a second clip he said xAI rewrote its training, reinforcement learning and inference software from scratch over six to eight months.
Deep dive
Elon Musk replied to a post from the account Fireside Alpha that Grok 4.8 is currently several different model variants in beta, adding that xAI is not yet sure which of the variants will go to wide release. His reply is the only first hand statement here, and it does not give a release date or any benchmark results.
The Fireside Alpha post he was answering shares a clip from Musk's interview at the New York Times DealBook Summit. In it, Musk says xAI is working on Grok 4.8 right now and that it is in the RL phase, meaning reinforcement learning. He says xAI does periodic soft releases under Grok Bot, so what users experience at times is various versions of Grok under the hood being tested for usefulness. He says the data on whether a version solved the user's problem is then used to improve Grok, which he called a good data flywheel.
In a second clip from the same interview, posted by Fireside Alpha, Musk says xAI rewrote its entire training, RL and inference stack from scratch rather than relying on JAX and SGLang, mainly in C and some C++, over six to eight months. He says he thinks xAI might now have the best training stack software in the world and that the results will show in the models that come out.
What to expect next
Watch for xAI to name which Grok 4.8 variant goes to wide release, and for an official model card or benchmark results. Independent testers will likely try to spot differences in Grok Bot answers that hint at which variant they are using. Musk's claim about xAI's rewritten training stack is his own and has not been independently checked, so results of the next release will be the first real evidence.
