The interesting part of Elon Musk‘s latest update on Grok is not the programming language. It is the admission folded inside it: the next model is late because the team building it threw away the open-source machinery it used to train and serve models, and wrote its own.

Read as a launch tease, the post is upbeat: training done, evaluations climbing, a soft rollout before the week is out. Read as a systems story, it says something bigger. xAI, now part of SpaceX, has decided that shared frameworks were holding it back, and it is betting this model, and the one after it, on owning every layer of software between its researchers and its GPUs.

“We have dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time,” Musk wrote on X, answering a user who asked when to expect Grok 4.8. JAX is the Python-based numerical library xAI had trained with; SGLang is an open-source serving framework for large language models. Pre-training is finished, he said, after “lot of cleanup was needed on the SFT data,” the supervised fine-tuning set. “Now, 4.8 is in RL and the evals are starting to climb.”

That line explains a gap Musk himself opened. In mid-September he introduced the model as “a 2.5T model trained with our new C++ software stack” that would “finish training this week and start RL.” Around the same time, as Analytics India Magazine reported, he called the reinforcement-learning side “still somewhat of a legacy pile of many different third-party software systems,” to be fixed “in due time.” By late October he was writing: “I know it’s taken longer than expected, but Grok 4.8 will not be a disappointment.” The fix turned out to be the delay. The pre-training code was already new; the part that teaches the model to behave, and the part that answers users, had to be rebuilt before the model could move on.

The “C++ rewrite” in the headlines is mostly C. Asked about the language, Musk replied: “C with a little C++, but not too much.” He says the stack descends from the software SpaceX engineers wrote for Starlink (“At SpaceX, we write our own headers and calculate our own time,” he added), and in September he said the pre-training code was written by that Starlink team and “was not written by AI,” because “AI is not yet good enough for extremely high-performance software (but it will be).”

The rationale has been public since spring. In late May, Musk said SpaceX had almost finished an in-house training stack in C that “exact-maps” to a cluster of 220,000 Nvidia GB300 chips, and put the potential speed-up over JAX for large training runs at “over an order of magnitude.” At the end of June he promised “truly massive gains” in about three months, once the whole training and inference stack was rewritten and “most software layers will be deleted completely.” Measured against that clock, the rewrite landed roughly six weeks late. None of the speed claims has been measured outside the company.

Owning the stack cuts both ways. Dropping JAX and SGLang frees xAI from other people’s abstractions and release schedules, and lets its engineers map the software directly onto one specific piece of hardware. The cost is that every bug in the training loop and in the serving path now belongs to xAI alone, in code no outside community has used before it reaches users.

Musk is also lowering expectations, hard. “Grok 4.8 will bear no resemblance to 4.7 (thankfully),” he wrote, a public put-down of the model xAI shipped on September 21. “It is not yet incredible, but, this time, the foundations are built right, so I expect it to evolve to be something special.” Less than an hour later, asked how 4.8 would compare with Anthropic‘s Opus 5.5, he answered: “Not as good. Opus 5.5 is a masterpiece.” His real promise sits further out: “Grok 5 should exceed Opus 5.5 and maybe that happens in 2 or 3 months.”

Everything else is still missing. xAI’s news page and developer documentation carry no entry for it, according to trackers that check them; there is no model card, no benchmark, no price and no API listing. The 2.5-trillion-parameter figure is Musk’s alone, as are the evaluation results he says are climbing. The soft rollout goes to Grok Bot users “later this week,” with no day given, and in October Musk said most of those requests would be handled by “a lightning-fast version of Grok 4.8.”

That makes this week’s rollout a test of the plumbing more than of the model. If the model holds up under real Grok Bot traffic, the rewrite becomes the platform Grok 5 is trained on. If it stumbles, there is no upstream framework left to blame.

Tags: Elon Musk, SpaceX, AI models, xAI, Grok