Elon Musk just told us xAI threw out JAX and SGLang and rewrote its entire RL and inference stack in C/C++. That is the interesting part of this story. The model itself, so far, is a promise.
On November 10, 2026, Musk said that Grok 4.8 has finished pretraining and is now in reinforcement learning. He added that “evals are starting to climb” and that a soft rollout to Grok Bot users would happen “later this week.” It is the first concrete schedule since his September tease of a 2.5-trillion-parameter model. But it is not a launch. There is no xAI newsroom post, no API price file, and no public benchmark table. The quotes come from wire relays, not a direct read of the original post, so trust the on-platform text if they differ.
Musk blamed the delay on two things: the stack rewrite, which “took some extra time,” and a lot of cleanup on the SFT data before RL could start. He also said 4.8 will “bear no resemblance” to 4.7, is “not yet incredible,” but has foundations “built right.” That is an unusually modest line for him, and a useful one. It suggests the team is still tuning, not coasting.
The infrastructure detail is what builders should care about. xAI’s original Grok announcement in 2023 described a stack on Kubernetes, Rust, and JAX. SGLang, meanwhile, is a popular open inference engine for agentic workloads and RL rollouts. Dropping both for in-house C/C++ means xAI now owns the whole path from rollout to serving. Plenty of teams debate SGLang or vLLM against custom code for last-mile latency. xAI just spent months of calendar time on the custom side. That doesn’t prove C++ wins for you, since few teams run anything like xAI’s Colossus cluster. And Musk published no throughput, utilization, or latency numbers against SGLang-served Grok 4.7. Treat it as engineering intent, not evidence.
The rollout target is also narrow. Grok Bot is xAI’s persistent, tool-using agent product, not the Grok.com chat app and not the API. Musk gave no cohort size, no region, and no tier rules. Nothing was promised for Grok.com defaults, Grok Build, cloud marketplaces, or Cursor’s model picker. Until a model string like grok-4.8 appears in an API changelog, assume 4.7 is still production everywhere outside that Bot slice.
The timing is a little awkward. Paying Bot subscribers complained on November 9 that quotas drained faster than expected, and Musk promised fixes. Widening a paid cohort to a model still in RL is risky, which probably explains the soft language. If you run agents on Bot, treat this as a canary. Log task success, token burn, and tool-error loops per session, and keep a rollback path to 4.7 in your own router config.
The SFT comment deserves attention too. SFT quality caps what RL can fix. Noisy data leads to reward hacking and strange refusals that cost GPU weeks to sand down. xAI hasn’t published an SFT card or any detail on datasets or safety filters. Compare that with Anthropic, which shipped Claude Fable 5.2 with documentation on November 9, one day before Musk’s post. Grok 4.7 launched in September with a news post and benchmarks, so xAI knows how to do this properly. It just hasn’t yet for 4.8.
So does it matter? Somewhat. The frontier is crowded this week, with Fable 5.2 out and Gemini 4 access chatter still circulating, and a Bot-only soft rollout won’t change anyone’s model choice. “Evals climbing” is an internal claim, not a leaderboard result. If you build on open models, nothing changes unless xAI releases weights or API pricing that undercuts Grok 4.7. Watch the API changelog, not the social thread, and wait for reproducible evals before you rank 4.8 against anything.
