Grok 4.8 is a 2.5T-parameter model, and it's still in the oven

Grok 4.8 will have 2.5 trillion parameters and is trained on xAI's new C++ software stack. Elon Musk says it "will finish training this week and start RL." So pretraining ends this week, which is not the same as a release. Reinforcement learning comes next, and that's the stage where Grok 4.7 got stuck.

The size jump is large. Grok 4.6, the current release, launched on 7 Aug with 1.5T parameters. Grok 4.7 is 2.1T, and 4.8 is about 67% bigger than 4.6. That puts it close to Moonshot's Kimi K3 at about 2.8T. Musk calls 4.8 "a noticeable improvement" over 4.7. He says 4.7 is "roughly on par with Opus 5.0" but still behind on multimodal.

The catch is that Grok 4.7 still isn't out. On 2 Sep Musk said it would "come out in 10 days," and that date has passed. On 11 Sep he said 4.7 needed "a few more days" of RL tuning, because it had been penalized too hard for long answers and was giving up on problems it could otherwise solve. docs.x.ai still lists nothing newer than grok-4.6. Musk has also said Grok 4.9 will be "Astra/Fable class" and that Grok 5 might be "better than anything."

Why it matters: xAI keeps adding parameters faster than it ships models, and on current form, "done training" still leaves weeks of RL before release.