Grok 4.8 is finally coming, Musk: testing will open to some Grok Bot users this week

Beating AI News Flash: Musk finally explained why Grok 4.8 has been delayed. The team abandoned the original JAX and SGLang software stack and rewrote the entire reinforcement learning and inference system from scratch in C/C++, which took considerable extra time.

On September 14, Musk revealed that Grok 4.8, with 2.5 trillion parameters, was about to complete pre-training. At the time, he admitted that pre-training had already switched to a self-developed C++ system, but reinforcement learning still relied on multiple patched-together legacy software components. Now that part has also been rewritten.

Currently, Grok 4.8 has completed pre-training, and the team has also put considerable effort into cleaning up supervised fine-tuning (SFT) data—the data used to teach the model to answer questions according to instructions. The model has now entered the reinforcement learning stage, and internal evaluation scores are starting to rise. Musk plans to open testing to some Grok Bot users this week.

Musk also bluntly criticized the previous generation: "Thank goodness, Grok 4.8 will be completely different from 4.7." But he also admitted that the new model is not yet impressive, but this time the underlying foundation has finally been properly built, and he expects greater improvements to follow.