Musk updated the progress of Grok 4.8: pre-training has been completed, now it has entered the reinforcement learning stage, and the evaluation scores have begun to go up. It will be released first to Bot users in grayscale later this week.
This time the delay took longer than expected. The reason he gave was that the entire bottom floor had to be replaced. xAI deprecated JAX and SGLang, and rewrote the entire reinforcement learning and reasoning stack from scratch in C/C++.
JAX is Google's numerical computing framework, and SGLang is a large model inference engine commonly used in the open source circle. Most teams will optimize on these off-the-shelf components, and completely throwing them away and rewriting them in C/C++ is a heavy engineering decision. The advantage is that each layer can improve its performance according to its own hardware and model structure. The cost is time, and for a period of time, you have to fix any problems yourself.
Musk was unusually restrained in his assessment of the new model. He said the 4.8 and 4.7 bear little resemblance and are "not amazing" yet, but the basics are right.
This can be heard from both sides. The optimistic reading is that the things that were overthrown and rebuilt have just been passed, the scores are still climbing, and there is room behind. A conservative reading is that other companies have released new flagships intensively in the past two weeks. If Grok 4.8 Grayscale is only "not amazing" when it comes out, it will not be good on the list.
As for why he chose to change the base at this time, he did not elaborate. A reasonable guess is cost. After the number of users of xAI increases, the cost of inference will be burned on a daily basis. Even if the self-developed inference stack only saves 10% or 20% of the computing power, it is still a lot of money to put it on a cluster of its size.
Grayscale will first give it to Bot users, that is, subscribers of the Grok Bot package. There is no timetable for when ordinary users will be able to use it.