10 Fascinating Facts About Elon Musk’s Grok 3 Model

Overview

Grok 3 is released on the dedicated website grok.com and as an independent iOS application. Access is currently free.

Model Variants

Grok 3 Base – 1.8 trillion‑parameter model delivering the highest accuracy.

Grok 3 Mini – 400 billion‑parameter version optimized for faster response.

Grok 3 Reasoning – chain‑of‑thought architecture for complex problem solving.

Grok 3 Mini Reasoning – balances speed with deeper reasoning capabilities.

Benchmark Positioning

During the launch event xAI removed OpenAI’s o3 inference model from the benchmark charts, implying that Grok 3 would outperform o3 if it were re‑inserted.

Performance Benchmarks

In the LMSYS chatbot arena Grok 3 achieved the top rank with an ELO score above 1400, surpassing GPT‑4o (1385) and Claude 3.5 Sonnet (1372). It recorded a 72 % preferred‑response rate, especially strong on physical‑simulation and creative‑coding challenges.

Key benchmark results:

AIME mathematics benchmark – 92 % accuracy.

GPQA science test – placed in the 94th percentile of human performance.

Supercomputer Infrastructure

Grok 3 runs on the custom‑built Colossus supercomputer in Memphis. The system started with 100 000 Nvidia H100 GPUs and was later expanded to 200 000 GPUs. Training consumed over 200 million GPU‑hours, enabling a 1.8 trillion‑parameter architecture—ten times larger than Grok 2.

Demonstrations

Live demos included a Mars‑Earth orbital transfer simulation with optimal Δv calculations, a hybrid Tetris‑Gem puzzle game, and a 3D‑printed model of a quantum‑computing component.

DeepSearch Feature

Grok 3 introduces DeepSearch, an AI‑assisted research tool that queries academic papers and social data. Example: when asked about the next Starship launch, DeepSearch cited over 127 sources, cross‑referenced orbital mechanics, and provided confidence‑weighted predictions.

iOS Application

The independent Grok iOS app is available in the Apple App Store for iPhone and iPad. Users can switch to the Grok 3 model via a dropdown menu in the chat window.