A small group of subscribers to the Grok chatbot has begun receiving access to a new version of the software, although a broad official release may not occur until late November or early December. The limited deployment serves as a testing phase to evaluate the model’s stability and performance before it becomes available to all users.
The Grok 4.8 model is built on a fundamentally different technical architecture than its predecessor. xAI engineers discarded the JAX and SGLang frameworks to rewrite the reinforcement learning and inference stacks from scratch using C and C++, a move Elon Musk said was influenced by the engineering philosophy used for the Starlink satellite network. While the model has completed pre-training and initial data cleanup, it remains in the reinforcement learning stage as evaluation scores begin to climb.
