Grok
Frontier reasoning models, trained on Colossus
Grok is xAI's family of large language models, trained end-to-end on the Colossus GPU cluster and deployed directly into X, Tesla vehicles, and Optimus. Rather than a single chatbot product, it's one model family embedded across every surface the group ships — the same reasoning engine answering a question on X, planning a route in a Tesla, or deciding a grip motion in Optimus's hand.
Trained on Colossus
A dedicated 230,000+ GPU cluster trains every generation of Grok, purpose-built rather than rented capacity from a third party.
Everywhere at once
The same model family answers questions on X, plans routes in a Tesla, and reasons through pick-and-place tasks in Optimus.
Real-time reasoning
Low-latency inference keeps Grok usable for live conversation, in-cabin voice commands, and robotic decision-making, not just batch queries.
Continuous retraining
Fleet and platform data feeds back into training runs, so the model improves from real-world deployment rather than static datasets alone.
One model, every surface
The X app, Tesla's in-car assistant, and Optimus's decision-making all call the same underlying Grok model family, just fine-tuned per surface — a bug fix or capability gain in one shows up in all three.
Built for scale from day one
Grok was trained on a cluster built specifically for it, not rented GPU time — giving xAI control over the full training pipeline as model sizes grow.
Model
Training
Access
01 / 0302 / 0303 / 03