Musk plans limited Grok 4.8 rollout after code rewrite

Elon Musk said Tuesday that Grok 4.8 will begin a limited rollout later this week. He said the company behind the chatbot had replaced its outside software frameworks with its own code, which delayed the model. "We have dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time," Musk wrote on X, according to Anadolu Agency. The company is now called SpaceX Super Intelligence and was formerly known as xAI.breakingthenews.net

The update comes as xAI tries to move users off Grok 4.7, which launched in September. Musk was openly critical of that model and gave no benchmark scores for the new one.

Where Grok 4.8 stands

Musk said pre-training is done and that "a lot of cleanup was needed on the SFT data to get that mostly right." SFT is supervised fine-tuning. He added that "4.8 is in RL and the evals are starting to climb," according to FourWeekMBA, which quoted the post as published at 09:35 UTC on November 10. RL refers to reinforcement learning, the stage where a model is trained against rewards. "We will do a soft rollout to Bot users later this week and see how it goes," Musk wrote. He did not say how large the rollout will be, whether API customers will get access, or what the model will cost, FourWeekMBA noted.fourweekmba.com

The Tesla CEO also took a swipe at the current model. "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," he wrote. Grok 4.7 launched on September 21.breakingthenews.net

A 57-day slip

The post came 57 days after Musk wrote on September 14 that Grok 4.8 "will finish training this week and start RL," FourWeekMBA calculated. In that earlier post, Musk described 4.8 as "a 2.5T model trained with our new C++ software stack." That meant pre-training had already moved to in-house code. The new post says the reinforcement learning and inference stages have now moved as well. Reinforcement learning had still relied on older patched software stacks at the time of the September post.lookonchain.com+1

The change is a break from how Grok was first built. xAI's November 2023 launch post said the company had built "a custom training and inference stack based on Kubernetes, Rust, and JAX," FourWeekMBA reported. JAX is an array-computing library from Google researchers. SGLang is an open-source framework for serving large language models.fourweekmba.com

Competitive timing

The soft launch comes days after rival releases. Anthropic announced Claude Fable 5.2 on November 9, and Google announced Gemini 4 Carbon on October 28, according to FourWeekMBA. It also comes after a week of complaints from Grok Bot users. On November 9, Musk promised overnight fixes after paid subscribers said their usage allowances were running out.fourweekmba.com

For now, Musk's word is the only evidence for both the stack rewrite and the climbing eval scores. Outside testing of Grok 4.8 will have to wait until users can try it.