Back to news
researchxAI2026-11-10

Musk said on November 10 that Grok 4.8 finished pre-training and entered reinforcement learning after the C/C++ stack rewrite

Musk said on November 10 that Grok 4.8 has finished pre-training and is in reinforcement learning with internal evaluation scores climbing. The team dropped JAX and SGLang and rewrote the entire reinforcement learning and inference stack in C and C++, with a soft rollout to Grok Bot users later this week.

On November 10, Elon Musk disclosed the progress of the xAI model Grok 4.8. The product was originally planned for release at the end of October and has been delayed by more than half a month because the team abandoned JAX and SGLang and rewrote the entire reinforcement learning post-training and inference stack from scratch in C and C++. Pre-training for Grok 4.8 is now complete, the team spent considerable effort cleaning up supervised fine-tuning data, the model has entered the reinforcement learning phase and internal evaluation scores are climbing, with a soft release to Grok Bot users later this week. The model has 2.5 trillion parameters, more than the 2.1 trillion of Grok 4.7, while the earlier 4.6 and 4.5 versions had 1.5 trillion. Musk said Grok 4.8 is completely different from 4.7 but not yet stunning, that this time the foundation has been correctly rebuilt, and that the team will not overfit to AI leaderboards for benchmark scores. He also conceded the model cannot reach or surpass Claude Opus 5.5, calling Opus 5.5 a masterpiece, and predicted only Grok 5, arriving in two to three months, would surpass it. Unlike the previous 4.7 which was delayed at least five times, Grok 4.8 is opening only to some Bot users this week rather than releasing broadly.

Grok 4.8xAI强化学习C/C++2.5 万亿参数延迟发布