SIAgentNews.com
← All news
Developing

Elon Musk announces Grok 4.8 with a complete C++ stack rewrite

xAI is replacing its primary software frameworks with a custom C/C++ stack to improve performance and scalability.

· First reported by The News International

4.8Grok model version

The short version

  • The engineering team abandoned JAX and SGLang for a custom C/C++ stack.
  • Pre-training is finished following a cleanup of supervised fine-tuning data.
  • Early user reports indicate reduced resource consumption after system resets.
  • New features include individual bot tracking and an upcoming iMessage integration.

Why it matters

By moving to a bare-metal stack, xAI aims for greater control over how the model processes information. This shift suggests a focus on efficiency and architectural strength for future growth.

A complete architectural shift

Elon Musk recently announced the upcoming release of Grok 4.8, a version of the AI model that undergoes a significant structural overhaul. The engineering team at xAI has decided to move away from the industry-standard frameworks that previously powered the system. Instead, they have opted to build a custom stack from the ground up.

The primary change involves the total abandonment of JAX and SGLang. These are common software libraries used to train and run large-scale machine learning models efficiently. By removing them, the team is moving toward a more direct way of interacting with computer hardware.

As reported by The News International, this transition required significant effort from the engineering team. Musk noted that the move to a custom C/C++ stack took extra time because it involves writing the core logic manually. This shift is intended to provide a stronger architecture for the model's future growth.

How the new stack works

The new architecture focuses on a bare-metal C/C++ stack for reinforcement learning and inference. Reinforcement learning is a method where an AI learns by trial and error to achieve specific goals. Inference is the process of the model actually generating a response when a user asks a question.

Writing these components in C and C++ allows the developers to have granular control over memory and processing power. This is often referred to as 'bare-metal' programming because it sits very close to the actual hardware. It avoids the overhead that can sometimes come with using high-level, pre-built libraries.

The team spent considerable time cleaning up the supervised fine-tuning data to ensure the model was prepared for this new foundation. Now that the pre-training phase is finished, the model is entering the reinforcement learning stage. Musk stated that the evaluations for the model are already starting to climb as a result of this work.

Early performance and features

Users who have interacted with the updated system have already begun to notice specific technical improvements. Some early reports highlight a reduction in resource consumption following system resets. This suggests that the new C++ stack may be more efficient in how it manages the underlying computing power required to run the AI.

Beyond the backend changes, the update includes several user-facing features and upgrades. These include the ability to track individual bots and the launch of an educational hub. This new hub will feature various AI programs designed for learning purposes.

The update also hints at expanded accessibility for the Grok ecosystem. For example, the team is working on upcoming compatibility with iMessage. These additions aim to make the model more integrated into the daily workflows and communication tools of its users.

What to expect next

The rollout for Grok 4.8 is scheduled to begin as a soft launch later this week. A soft rollout means the update will be released to a limited group of Grok Bot users first. This allows the team to monitor the performance of the new C/C++ stack in a live environment before a wider release.

Musk emphasized that this version will have no similarity to the previous version, 4.7. This indicates that the changes are not just minor tweaks but a fundamental shift in how the model operates. The goal is to create a more robust system that can handle more complex tasks as the technology evolves.

The engineering team will likely be monitoring the feedback from these early users closely. They will be looking for any issues related to the new inference stack or the reinforcement learning process. The success of this transition will determine how the model scales in the coming months.

Technical complexity and impact

The move to a custom stack is a significant undertaking for any AI company. Many developers rely on established libraries to handle the heavy lifting of model deployment. By choosing to write their own C headers without using other libraries, the xAI team is taking a much more difficult and manual path.

This level of engineering is often seen as a way to squeeze out maximum performance from hardware. It allows for optimizations that are not possible when using generalized software. For the user, this might translate to faster response times or the ability to run more complex reasoning tasks.

The community has already reacted to the scale of this project with a mix of surprise and interest. Some users have pointed out the complexity involved in building such a system from scratch. It represents a bold bet on the idea that custom engineering can provide a competitive edge in the rapidly evolving AI landscape.

What to watch

  • How the C++ rewrite affects response speed compared to the old JAX stack.
  • The final results of the reinforcement learning evaluations as the soft rollout expands.

Try it

Check the Grok Bot later this week to see if you can access the new 4.8 features.

Sources and links

First reported by The News International. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news