Elon Musk has revealed new details about the software infrastructure being used to train Grok 4.8, saying the upcoming 2.5-trillion-parameter AI model is being pre-trained using an internally developed C and C++ stack written primarily by human engineers.
The disclosure came as Musk shared updates on the next generation of Grok models. According to him, Grok 4.8 is expected to complete its current training phase this week before moving into reinforcement learning, or RL. The model has not yet been released, and xAI has not published a technical report or model card detailing its architecture or performance.
Musk said the new pre-training software was developed primarily by the Starlink software team. He described the system as an internally built C/C++ stack and said it was written by humans rather than generated by artificial intelligence.
Responding to a question about whether AI had written the training software, Musk said AI is not yet capable enough for the level of high-performance software required by the company, although he expects that to change.
The approach marks a shift from xAI's earlier training infrastructure. Grok-1, released in 2024, was trained using a custom stack built on JAX and Rust. Musk has previously discussed plans to move more of the company's AI training and inference infrastructure to internally developed C and C++ software.
The strategy is aimed at reducing software layers between AI workloads and computing hardware. Musk has previously said that developing software closer to the underlying hardware could improve training efficiency, although the company has not published independent benchmarks demonstrating the performance gains of the new stack.
The reinforcement learning infrastructure has not undergone the same transition. Musk said the company's RL stack still consists of multiple third-party software systems and described it as a legacy setup that remains to be rebuilt.
Grok 4.8 is expected to have 2.5 trillion parameters, according to Musk. However, parameter count alone does not determine a model's performance, which also depends on factors including training data, compute, optimisation techniques and post-training.
The update comes while xAI continues work on Grok 4.7, which has not yet been publicly released. Musk has indicated that Grok 4.8 should improve upon that model, while also discussing Grok 4.9 and Grok 5 as later additions to the company's roadmap.
No public release date has been announced for Grok 4.8.
Disclaimer: This article may include information derived from interviews, press releases, public statements, research, company communications and other publicly available or third-party sources. Such material may be summarised, paraphrased or contextualised for journalistic and editorial purposes. All rights in third-party content remain with their respective owners.