Breaking

Elon Musk Says Grok 4.8 Is a 2.5 Trillion-Parameter Model on a New C++ Stack

September 14, 2026 at 05:33 EDT

Foundation Models
Infra & Chips

Elon Musk said Grok 4.8 will be a 2.5 trillion-parameter model trained on SpaceXAI's new C++ software stack, with training due to finish this week before reinforcement learning begins. The company has not published a model page, API identifier, price, or launch date, and its public news feed still lists Grok 4.6 as the latest released flagship.

September 14, 2026 · SpaceXAI

Grok 4.8 named as a 2.5T model

Elon Musk said Grok 4.8 is a 2.5 trillion-parameter model trained on a new C++ software stack. Pre-training is due to finish this week, then reinforcement learning begins. No ship date or official spec sheet has been published.

2.5T

Stated Grok 4.8 size

~67%

Larger than Grok 4.6 (1.5T)

This week

Pre-training, then RL

Stated size ladder

Musk's figures, in trillions of parameters. SpaceXAI has not published a count for any Grok 4 model.

1.5T

Grok 4.6 shipped

2.1T

Grok 4.7 delayed

2.5T

Grok 4.8 in training

Takeaway: 4.8 is stated at about two-thirds larger than the 1.5T model now in production.

What comes next

Pre-training ends

This week

→

Reinforcement learning

No end date given

→

Public release

Not announced

Interest

First numbered Grok said to be trained on the new C++ stack, at a stated 2.5T — about 67% above production 4.6.

Caution

Grok 4.7 is still delayed in RL. 4.8 has no ship date, price, or independent evals; finishing training is not a launch.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.