10 Things You Should Know About Grok 3 AI Model

Elon Musk and xAI's new Grok 3 model has just been released. Grok 3 is 10 to 15 times more powerful than Grok 2. It runs on the Colossus supercomputer, using 100,000 Nvidia H100 GPUs and racking up 200 million GPU-hours for training. With that kind of power, Grok 3 can handle massive datasets fast and accurately, pushing AI computing to a whole new level.

Here are 10 interesting facts about the recently launched Grok 3 model:

1. New Grok website launched
2. Grok 3 has four model variations
3. Selective benchmarking?
4. Musk claims Grok 3 is the smartest AI on Earth
5. Grok 3 runs on Colossus Supercomputer
6. Grok 3 tops Chatbot Arena benchmark
7. Grok 3 beats GPT-4o in key benchmarks
8. Grok 3 can build games and simulators
9. Grok 3 challenges DeepSeek with DeepSearch
10. xAI launches a standalone Grok app for iOS

1. New Website Launch

Grok 3 will soon be available on the new website, grok.com. This dedicated platform will let users explore and interact with the AI model in an easier way.

2. Four Model Variations

Grok 3 comes in four distinct variations:

1. Grok 3 Base: The full-scale 1.8 trillion parameter model for maximum accuracy.
2. Grok 3 Mini: A 400 billion parameter version that is faster in responding.
3. Grok 3 Reasoning: Uses a chain-of-thought architecture to solve complex problems.
4. Grok 3 Mini Reasoning: Balances speed with analytical depth.

Each variant is designed to cater to different needs and use cases.

3. Selective Benchmarking?

During the Grok 3 announcement livesteam, xAI omitted OpenAI’s o3 reasoning model from the benchmark charts.

If o3 is inserted in the benchmark, it will overthrow the Grok 3 reasoning model.

This move hints at strategic positioning, suggesting that Grok 3 is ahead of some of its competitors in certain areas.

4. Smartest AI on Earth

Elon Musk has called Grok 3 the “smartest AI on earth.”

This bold claim comes after the model’s impressive performance in multiple evaluations and live demonstrations.

Is it, though?

5. The Colossus Supercomputer

Grok 3 runs on the Colossus supercomputer, built in Memphis specifically for its development.

• Hardware Details: It initially used 100,000 Nvidia H100 GPUs, later doubling to 200,000 units.
• Training Effort: The training process took over 200 million GPU hours.

This incredible hardware setup enables Grok 3’s huge 1.8 trillion parameter architecture—10 times larger than its predecessor, Grok 2.

6. Chatbot Arena Champion

Grok 3 ranks number one on LMSYS’ Chatbot Arena. It is the first AI model ever to score over 1400 ELO.

This achievement shows its dominance in public tests, where it outperformed other leading AI systems like GPT-4o.

7. Grok 3 Beats GPT-4o in Key Benchmarks

In side-by-side tests, Grok 3 outperformed GPT-4o on several important benchmarks:

• AIME Mathematics Benchmark: Grok 3 scored 92% accuracy.
• GPQA Science Test: Grok 3 reached the 94th percentile of human performance.

These tests cover everything from math to PhD-level science, proving Grok 3’s advanced capabilities.

8. Rapid Improvement

Musk highlighted the quick progress of Grok 3 by stating, “Grok is ready to go to college.” This remark underlines how much the model has advanced since Grok 2, reflecting a steep learning curve and fast evolution.

Grok 3 is also really good at creating games and simulators. During demonstrations, Grok 3 showed off its creative side:

• It designed a Mars-Earth orbital transfer simulation with optimal Δv calculations.
• It created a hybrid game that mixes elements of Tetris and Bejeweled.
• It even generated 3D printable models for theoretical quantum computing parts.

Musk noted that these outputs show “the beginning of artificial creativity,” a sign of Grok 3’s potential to revolutionize both practical tasks and creative projects.

9. DeepSearch: Next-Generation Information Retrieval

Grok 3 introduces DeepSearch, a new feature for AI-assisted research.

• How It Works: It searches through academic papers and social data to find relevant information. If you heard about ChatGPT’s Deep Research or DeepSeek R-1 with Deep Research, Grok 3’s own DeepSearch is expected to work the same way.
• Practical Example: When asked about the next Starship launch, it identified over 127 sources, cross-referenced orbital mechanics, and even provided a confidence-weighted prediction.

This blend of neural search and symbolic reasoning makes DeepSearch a unique tool in the AI toolkit.

10. xAI Launches A Standalone Grok App for iOS

Along with the announcement of Grok 3, xAI launched a brand-new standalone Grok iOS app. You can now download the app on Apple’s App Store for iPhone and iPad.

To switch to the Grok 3 model, open a chat instance and set the model to Grok 3 in the upper portion dropdown menu.

‍