xAI announces Grok 4.8 completion of main training run and a 2.5T-parameter rewrite

xAI has announced Grok 4.8 as a 2.5T-parameter model trained on a new C++ stack and reports it will move to reinforcement learning; no official release materials or independent benchmarks have been published yet.

In this brief: 3 sections 2 min read

- • Elon Musk said Grok 4.8 completed main pretraining and will enter RL phase.
- • Reported size: ~2.5 trillion parameters; predecessor Grok 4.7 still unreleased.
- • xAI built a new C++ training stack in‑house per the report.
- • The report highlights absence of an official benchmark card or API/pricing table for Grok 4.8.
- • External verification (benchmark scores, release artifacts) is not yet available.
- • Author flags reliance on Musk’s social posts rather than published data.
- • A 2.5T model and custom C++ stack imply significant infrastructure and potential competitive differentiation.
- • Without benchmark cards or API release, developers and customers lack actionable details for adoption or comparison.
- • Report recommends watching official benchmark scores and release materials as proof points.

Read full analysis on winzheng.com

Useful?