Introducing Grok 4 and reinforcement learning scaling at xAI
xAI has released Grok 4, utilizing a massive 200,000 GPU cluster to scale up reinforcement learning training.
2 min readTuesday, 27 October
xAI has released Grok 4, accompanied by a new subscription option called the SuperGrok Heavy tier that grants access to Grok 4 Heavy.
Key facts
- GPU Cluster Size
- 200,000 GPUs
- Compute Efficiency Gain
- 6x
Scaling Reinforcement Learning
The development of Grok 4 builds on previous scaling work from Grok 3 and Grok 3 Reasoning. To train Grok 4, xAI utilized Colossus, a cluster containing 200,000 graphics processing units, which are specialized computer chips designed to handle heavy workloads. This setup ran reinforcement learning training—a method where artificial intelligence models learn by trial and error—at pretraining scale.
Source: Original announcement
Worth a read?
Comments