Musk draws another pie in the sky: 2.5T-parameter Grok4.8, pre-training to end this week and shift to RL.

Beating AI News Flash: Before Grok 4.7 is even released, Musk is already teasing Grok 4.8. He claims 4.8 is a 2.5T parameter model, trained with SpaceXAI's new C++ software stack, with pre-training set to finish this week, followed by reinforcement learning.

This C++ stack is also a pie Musk has repeatedly drawn before. At the end of June, he said he wanted to substantially rewrite Grok's training and inference software in C/C++ within about three months, remove a large number of intermediate layers, and make dedicated optimizations for Nvidia's GB300, expecting "huge improvements" in about three months.

However, Grok 4.7 was just delayed due to RL issues. Musk previously said it would be released around September 12, but near launch he admitted the model gives up too early on hard problems and is not strict enough when checking answers. Before the last pie is even out of the oven, the next one is already in.