Musk says 2.5T-parameter Grok 4.8, trained on new C++ stack, finishes training this week
inductionheads · x · 2026-09-14
Elon Musk revealed that Grok 4.8 is a 2.5T-parameter model trained with xAI's new C++ software stack. Training finishes this week, after which the model will enter the RL stage. If accurate, it would be among the largest known models by parameter count, and the C++ training stack marks an unusual technical choice.
More from Models
- Next Token Podcast Ep.2: GPT-6 Astra, DeepSeek V4.1 Flash and iPhone Duo — op7418 · 2026-09-14
- Three weird image prompts — pelican on a bike, Wukong flying — to detect model dumbing — lxfater · 2026-09-14
- "Imagine paying $200 a month": user can't get one light-mode question through — Beeegbong · 2026-09-14
- Most accurate multiclass brain segmentation model on Brainchop hits 0.9 Dice, now open for testing — PlisSergey · 2026-09-14
- Claude Computer Use is back, says developer who mourned its decline in favor of Claude Code — voooooogel · 2026-09-14
- DeepSeek V4.1 Flash takes first on Artificial Analysis' new v4.3 benchmark, dethroning Astra — Randomdotmath · 2026-09-14