Musk: Grok 4.8 ditches JAX/SGLang, RL and inference stack rewritten from scratch in C/C++
r0ck3t23 · x · 2026-11-11
Elon Musk confirmed Grok 4.8 was not built on the old system but by tearing it out: the team dropped JAX and SGLang and rewrote the entire reinforcement learning and inference stack from scratch in C and C++.
Key points:
- Pre-training is finished; SFT data needed a major cleanup before it was mostly right
- The model is now in RL, with evals starting to climb
- Musk says 4.8 will bear no resemblance to 4.7 ("thankfully") and admits it is "not incredible yet," but expects the rebuilt foundations to let it evolve into something special
- A soft rollout to Grok Bot users starts later this week
The author argues most labs ship a bigger number on the same plumbing; rewriting the stack costs months upfront but makes every subsequent model train faster and run cheaper.
More from Infra
- Microsoft & Tsinghua: 60% of GPU kernel library failures in LLM systems come from build and environment, not kernel code — rohanpaul_ai · 2026-11-11
- "Save water, run local models": dev's one-liner pokes at AI's water footprint — JeremyCMorgan · 2026-11-11
- First SGLang Summit draws 2,000+ attendees as the project grows from under 10% share to ecosystem core — ying11231 · 2026-11-11
- Protecting SSH Keys with macOS Secure Enclave: A Hands-On Guide — js2 · 2026-11-11
- Neon ships branch-aware object storage rollback, 24 function templates, CJK search — samgoodwin89 · 2026-11-11
- Anthropic's amended S-1: $170B+ Microsoft/Google compute deals, IPO at up to $1.4T valuation — rohanpaul_ai · 2026-11-11