Musk says Grok 4.8 pretraining is complete and it will be rolled out to a small group of Bot users this week
xAI founder Elon Musk wrote on X on November 10, 2026, in response to a user’s question, that the team had abandoned JAX/SGLang and rewritten the entire reinforcement learning and inference system from scratch in C/C++, which caused extra time to be spent. He said pretraining has been completed, and the supervised fine-tuning data have also been heavily cleaned up; Grok 4.8 is currently in the reinforcement learning stage, and evaluation scores have started to rise. xAI plans to first roll it out to a small group of Bot users later this week to observe the results.
Musk said Grok 4.8 is “completely different” from 4.7 and is currently “not yet that good,” but “the underlying foundation is now in the right place this time,” and he expects it to gradually evolve into a special model.
Original source:
Title: We have dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++... Publisher: Elon Musk (@elonmusk) Publication date: 2026-11-10 URL: https://x.com/elonmusk/status/2120088070443655190 Source nature: original post by the person involved Verification status: The full original post and context were actually read for this report. Verification date: 2026-11-10. Supporting content: Supports all key facts in the body, including the technology stack rewrite, completion of pretraining, cleanup of SFT data, the current RL stage and rising evaluation scores, the small rollout to Bot users later this week, and the description of how it differs from 4.7.
ABAB AI Insight
Since xAI was founded, Elon Musk has repeatedly disclosed progress on Grok iterations. On September 14, 2026, he said on X that Grok 4.8 is a 2.5T-parameter model and will use a new C++ software stack to complete pretraining before entering RL. This update confirms that pretraining has ended and the model has entered RL, while also making clear that the entire RL and inference stack was rewritten after abandoning JAX/SGLang, explaining the technical reason for the earlier delay.
In terms of the business path, xAI has shifted its technology stack from relying on existing frameworks to a fully self-developed C/C++ implementation in order to gain lower-level hardware control and efficiency. After large-scale cleanup of SFT data, the model entered RL and evaluation scores began to rise, indicating that the model has entered a period of observable improvement in alignment and task performance. A small rollout to Bot users is a typical phased validation approach, first collecting real-world usage feedback before expanding the rollout.
Compared with the earlier Grok 4.7, Musk made it clear that 4.8 is “completely different,” and also reflected on some issues with 4.7, such as giving up difficult tasks too early. This is similar to the approach taken by labs such as OpenAI and Anthropic, which improve stability by rewriting inference stacks or intensifying data cleaning during model iterations, but xAI chose to rewrite the stack from the ground up in a lower-level language, making it a more aggressive technical route.
This progress may give xAI differentiated advantages in inference and RL efficiency. If the C/C++ rewrite brings actual speed or cost improvements, it will strengthen its competitiveness in high-concurrency Bot scenarios; however, Musk himself admitted that it is currently “not yet that good,” and the final effect still depends on subsequent RL optimization and feedback from the phased testing.
ABAB News · Cognitive Law
A rewrite of the underlying technology stack can explain the delay in progress, but once completed, the pace of subsequent evaluation gains may accelerate.
When the person involved proactively admits that the model is “not yet that good,” it often provides more informational value than mere promotion.