Grok is a family of large language models developed by xAI, the AI company founded by Elon Musk in July 2023. Grok-1 (released open-source in March 2024 at 314B MoE parameters) was notable for being the largest openly released model at the time. Grok-2 and Grok-3 added real-time X (Twitter) data access, image generation, and reasoning capabilities. Grok-3 (February 2025) scored highest on several benchmarks at release and is integrated directly into X Premium.
Architecture: 314B MoE and what makes Grok different
| Model | Params | Architecture | Open? | Key feature |
|---|---|---|---|---|
| Grok-1 | 314B | MoE (8 experts, top-2) | Yes (Apache 2.0) | Largest open model at release |
| Grok-1.5 | Undisclosed | MoE | No | 128K context window |
| Grok-2 | Undisclosed | MoE | No | Real-time X data, image gen |
| Grok-3 | Undisclosed | MoE | No | 200K H100 training, reasoning |
Grok's defining differentiator is real-time access to X (Twitter) data — it can reference tweets, trending topics, and live news in responses. This makes it uniquely useful for questions about current events, breaking news, and public sentiment analysis. However, this also means Grok's outputs reflect the biases and misinformation present in X's ecosystem.
Colossus: xAI's training cluster
Grok-3 was trained on Colossus — a cluster xAI claims contains 200,000 Nvidia H100 GPUs built in Memphis, Tennessee in just 122 days. By comparison, most frontier labs have 10,000–50,000 H100s. If accurate, this makes xAI one of the best-resourced AI training operations in the world.
Grok vs GPT-4o vs Claude: practical differences
- Grok: Best for real-time information, current events, Twitter/social media analysis
- GPT-4o: Best for coding, broad task coverage, most mature ecosystem
- Claude: Best for long documents, nuanced writing, safety-sensitive applications
- Gemini: Best for Google Workspace integration, video understanding, Android apps
- All four are competitive on standard benchmarks — task fit matters more than benchmark scores