Skip to main content
LLMgram · AI News · 2026-11-10

Grok 4.8 enters RL after xAI rebuilds inference and RL stack in C and C++

Grok 4.8 enters RL after xAI rebuilds inference and RL stack in C and C++

xAI has moved Grok 4.8 into reinforcement learning after Elon Musk disclosed a ground-up rebuild of the company’s RL and inference stack. According to Musk’s reply on X, xAI abandoned JAX and SGLang and reimplemented the full stack in C and C++, a shift he said added schedule time but replaced the prior Python-oriented tooling chain. He stated that pre-training is complete, while supervised fine-tuning required extensive data cleanup before the pipeline was mostly right. Musk added that Grok 4.8 is now in the RL phase and that evaluations are beginning. The post offers a rare look at infrastructure choices behind the next Grok revision, though it does not specify benchmark results, release timing, or independent verification beyond Musk’s account.

Sources

Grok 4.8 enters RL after xAI rebuilds inference and RL stack in C and C++

Grok 4.8 enters RL after xAI rebuilds inference and RL stack in C and C++

In a reply on X, Elon Musk said xAI dropped JAX and SGLang and rewrote the RL and inference stack from scratch in C and C++. He said pre-training is done, SFT data needed substantial cleanup, and Grok 4.8 is now in RL as evaluations begin.

Key takeaway

xAI is betting Grok 4.8 on a from-scratch C/C++ RL and inference stack after dropping JAX and SGLang, with RL training and evals now underway.

What happened

In a reply on X, Elon Musk said xAI dropped JAX and SGLang and rewrote the reinforcement-learning and inference stack from scratch in C and C++. He said the rewrite took extra time relative to continuing on the prior stack.

Musk said pre-training is finished, that supervised fine-tuning data needed substantial cleanup, and that Grok 4.8 is now in reinforcement learning as evaluations begin.

Evidence

  • xAI abandoned JAX and SGLang and rebuilt the RL and inference stack in C and C++.

    X · attributed

    We have dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time.

  • Pre-training is complete and SFT data required substantial cleanup.

    X · attributed

    Training (aka pre-training) is done. Lot of cleanup was needed on the SFT data to get that mostly right.

  • Grok 4.8 is in RL and evaluations are starting.

    X · attributed

    Now, 4.8 is in RL and the evals are starting to

Why it matters

A full-stack move to C and C++ for RL and inference signals xAI is optimizing for control and performance at the serving and training boundary, not just model weights.

Limits and uncertainties

Musk’s post does not publish eval metrics, RL objective details, or a firm Grok 4.8 ship date.

A separate X snippet mentions a soft Grok Bot rollout this week but that claim is not attributed in the primary Musk status and lacks a distinct source URL in the packet.

Practical implications

Teams comparing stacks should weigh migration cost from JAX/SGLang-style pipelines against latency and operability gains from custom C/C++ inference and RL loops.

Builders tracking Grok should expect the next public signal to be eval outcomes or user-facing rollouts rather than further stack anecdotes.

What to watch

Whether xAI publishes Grok 4.8 evaluation results or benchmarks after RL completes.

Any official announcement of wider Grok Bot or API access tied to the 4.8 line.

Sources

LLMgram editorial selection and synthesis · @llmgram. LLMgram is not the original publisher of this information.
Continue on LLMgram: Open in AI Signal →
Original reporting: @cradrrat We have dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time. Training (aka pre-training) is done. Lot of