Taking longer, but will not disappoint
Elon Musk, who leads SpaceXAI (formerly xAI), a unit of SpaceX, acknowledged in an X post on Oct. 30 (the afternoon of Oct. 30 Japan time) that the next model, "Grok 4.8," is taking longer than expected, and said it will not disappoint. He also said the model has not been tuned toward benchmark scores and that he does not care about the scores, adding that the focus is on being useful in practice. He did not give a release date.
Training of "2.5 trillion model" was slated to finish in September
On Sept. 14, Musk posted that a 2.5 trillion-parameter model trained with new C++ software would finish pretraining that week and move on to reinforcement learning. This is believed to be what will become Grok 4.8. After pretraining, finishing work such as reinforcement learning is still needed.
On Oct. 7, he also said a fast version of Grok 4.8 would be used for everyday requests in the AI agent "Grok Bot." As of Oct. 30, xAI's official documentation does not mention Grok 4.8, and it is not available through the API.
The previous version was also delayed
The current Grok 4.7 was released on Sept. 21. On Sept. 11, Musk explained the reason for its delay, saying that reinforcement learning may have penalized answer length too harshly, causing the model to give up early on difficult problems.
APIYI cautioned that the remark about not emphasizing benchmarks does not mean Grok 4.8 performs poorly.
