On September 14, Elon Musk announced that the Grok 4.8 AI model, featuring 2.5 trillion parameters, is completing its initial training phase. The model is now transitioning into the reinforcement learning stage, which is intended to improve instruction following and reasoning capabilities.
This release follows the launch of Grok 4.6 on August 12, which contained 1.5 trillion parameters. The company bypassed the planned 2.1 trillion-parameter Grok 4.7 version due to development delays, opting to prioritize the more advanced 2.5 trillion-parameter iteration.
xAI utilizes its Colossus 2 supercluster for these training tasks, which is described as the world's first gigawatt-scale training cluster. The company is also developing foundation models that may reach up to 10 trillion parameters in the future.
The training process incorporates proprietary data from SpaceX operations and Cursor engineering workflows. Musk attributes the performance improvements in new releases to three factors: increased scale, higher data quality, and architectural design refinements.
xAI maintains a monthly release schedule, which is significantly faster than the quarterly or semi-annual cycles typical of other major AI laboratories. This high-velocity approach is supported by the company's specialized infrastructure and proprietary data access.
