Grok-3 Unveiled by xAI

Grok-3, developed by Elon Musk’s xAI, was unveiled on Monday, showcasing a massive computing infrastructure indicating bigger ambitions.

The announcement spotlighted raw computational power, benchmarks, and future features, though many demonstrations resembled those from other AI companies.

The highlight was “Colossus,” a 200,000 GPU cluster powering Grok-3’s training, which underwent 122 days of initial training on 100,000 GPUs, followed by 92 days scaling to full capacity. xAI developers claimed building this infrastructure was more challenging than developing the AI model.

xAI plans to build an even more powerful cluster aiming for five times the current capacity, potentially the world’s most powerful GPU cluster.

In terms of performance, Grok-3 demonstrates solid results on standard AI benchmarks, topping math (AIME), science (GPOA), and coding (LCB) tests without advanced features. It showed promise in blind tests, with a mysterious variant called “Chocolate” performing well in user preference.

A specialized variant, “Reasoning Beta,” utilizes internal chain of thought, achieving a higher math score of 93% on AIME 2025 benchmarks.

Despite some impressive live demonstrations, many tasks mirrored capabilities of established competitors like ChatGPT and Google’s Gemini, leading to perceptions of Grok-3 playing catch-up.

New features include DeepSearch, a web-scraping research agent, and upcoming voice interactions—unlike simple text-to-speech, it claims to offer expressive speech capabilities.

Grok-3 is slowly rolling out with API access expected soon. Early users and experts have praised its capabilities, comparing them favorably with existing market leaders like OpenAI and DeepMind.

The long-term test will be the performance of its promised features as it hits the market, while OpenAI is set to launch GPT-4.5 next.

Edited by Sebastian Sinclair