[Musk Teases Again: 2.5T Parameter Grok 4.8, Pretraining to RL Transition by End of Week] Beating AI Newsflash: Grok 4.7 hasn’t been released yet, but Musk is already teasing Grok 4.8. He claims that 4.8 is a model with 2.5T parameters, trained using SpaceXAI’s new C++ software stack. Pretraining is expected to conclude this week, followed by reinforcement learning. This C++ stack is another one of Musk’s repeatedly teased promises. Back in late June, he stated that within approximately three months, Grok’s training and inference software would undergo a major rewrite into C/C++, removing numerous intermediate layers and optimizing specifically for NVIDIA’s GB300, with expectations of delivering “significant improvements” in about three months. However, Grok 4.7 was recently delayed due to RL issues. Musk previously mentioned it would be released around September 12, but as the launch approached, he admitted the model tended to give up too early on difficult problems and wasn’t rigorous enough in checking answers. The previous promise hasn’t been delivered yet, and the next one is already in the oven. [Original Link]