Grok-3', the 'World's Smartest' AI Model, Released
Key Points
- 1xAI officially launched Grok-3, its next-generation AI model trained on the Colossus data center, claiming superior performance over OpenAI's GPT-4o and other leading models across various benchmarks.
- 2Grok-3 is a multimodal large language model available in four versions, including a specialized "Reasoning" model, and introduces "DeepSearch," an advanced AI inference agent capable of contextually analyzing internet search results.
- 3This release intensifies competition in the AI market, with Grok-3 available via X's Premium+ subscription tiers, and xAI plans to extend its capabilities to a voice assistant and open-source earlier Grok models.
xAI, led by Elon Musk, officially launched its next-generation AI model, Grok-3, on February 17, 2025 (local time) via an X (Twitter) streaming event. This announcement significantly intensifies competition within the AI industry, particularly with OpenAI, Google DeepMind, and Meta. Alongside Grok-3, xAI also unveiled 'DeepSearch,' an AI inference agent akin to OpenAI's 'Deep Research.'
Grok-3 is positioned as a leading AI model, having been trained at 'Colossus,' described as the world's largest AI data center, utilizing 100,000 GPUs located in Memphis, USA. It is presented as outperforming OpenAI's latest models, GPT-4o and 'o3-mini-hi.'
The core methodology of Grok-3 emphasizes its Large Multimodal Model (LMM) design, capable of not only advanced text generation but also image processing, enhancing its understanding and generation capabilities across various input data types. The model is released in four distinct versions:
- Grok-3 Mini: A compact model designed for lightweight operations and rapid response times.
- Grok-3 Reasoning: A specialized model offering advanced inference capabilities. It incorporates a 'Big Brain' mode that allocates additional computational resources for addressing complex queries and delivering in-depth answers, enabling the AI to "think."
- Grok-3 Mini Reasoning: A lightweight version of the reasoning model, prioritizing computational efficiency.
- Super Grok: A premium offering that bundles the reasoning model, DeepSearch, and unlimited image generation.
A notable new feature is DeepSearch, which allows the AI to perform internet searches on behalf of users, comprehensively analyze the retrieved content, understand context, and generate sophisticated, highly accurate answers through deep inference. This functionality is slated for integration into xAI's enterprise API with Grok-3 within weeks. Unlike conventional web search, DeepSearch's distinction lies in its AI-driven analysis and contextual understanding.
Grok-3's performance is substantiated by impressive benchmark results that surpass competing models:
- AIME 2025 (Mathematics): Grok-3 scored 52 points (vs. GPT-4o's 39 points). Specifically, the Grok-3 Reasoning model achieved 93 points, outperforming OpenAI's o1 and o3-mini-hi.
- GPQA (PhD-level scientific knowledge): Grok-3 scored 75 points (vs. GPT-4o's 65 points).
- Coding Ability Test: Grok-3 scored 57 points (vs. DeepSeek-V3's 40 points).
- IM Arena Leaderboard (user preference-based evaluation): Grok-3 garnered higher scores than GPT-4o and the latest Gemini 2.0.
In terms of service provision and pricing, Grok-3 and Grok-3 Mini are available through X's 'Premium+' subscription for \$22 per month. The 'Super Grok' service, priced at \$30 per month, includes the Reasoning model, DeepSearch, and unlimited image generation capabilities.
Future plans include expanding Grok-3 into an AI voice assistant, with voice mode integration into the Grok App expected within a week. Furthermore, Musk announced plans to open-source the previous Grok-1 and Grok-2 models, emphasizing transparency and accessibility in AI research while strategically differentiating xAI in the competitive landscape.