News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
Grok 3

Grok 3

xAI's reasoning-focused language model, released in beta in 2025

Grok 3 is a large language model from xAI, unveiled in February 2025 as an early preview. The company says it was trained on the Colossus supercluster with ten times the compute of earlier models, and it pairs broad pretraining knowledge with reasoning that can run for minutes at a time.

Share on WhatsApp

  • Trained on xAI's Colossus supercluster with 10x the compute of earlier leading models
  • Think mode can spend seconds to minutes working through a problem
  • xAI reports 93.3% on the 2025 AIME at its highest test-time compute setting
  • Elo of 1402 reported in Chatbot Arena
  • Grok 3 mini targets lower-cost reasoning for STEM tasks

What it is

Grok 3 is a language model built by xAI, shown publicly as a beta in February 2025. The company describes it as its strongest model to date, combining wide-ranging knowledge from pretraining with deliberate reasoning. It is aimed at reasoning, mathematics, coding, world knowledge and following instructions. A smaller sibling, Grok 3 mini, targets cheaper reasoning on STEM problems that lean less on general knowledge. Both were still being trained at the time of the announcement, and xAI said they would change quickly as user feedback arrived.

How it came about

xAI presented Grok 3 as an early preview alongside Grok 3 mini. The pair were announced as beta reasoning models, Grok 3 (Think) and Grok 3 mini (Think). Training took place on the Colossus supercluster. At the time of the announcement both systems were still in training, and the company said a rollout to users would follow within days, together with a first look at the reasoning features.

Inside the model

The reasoning behaviour was shaped by reinforcement learning applied at very large scale, which xAI says taught the model to refine its chain of thought in a data-efficient way. In the Think mode it can backtrack to fix mistakes, trim unnecessary steps and draw on what it learned during pretraining. It may spend a few seconds or several minutes on a problem, try more than one route, check its own answer and aim to fit the exact requirements. Training ran on the Colossus supercluster with ten times the compute of earlier leading models.

What it can do

With the highest test-time compute setting (cons@64), Grok 3 (Think) scored 93.3% on the 2025 American Invitational Mathematics Examination, a paper published seven days before the test. xAI also reports 84.6% on the graduate-level GPQA science set and 79.4% on LiveCodeBench for code work. Grok 3 mini reached 95.8% on AIME 2024 and 80.4% on LiveCodeBench. The company reports an Elo of 1402 in Chatbot Arena. With reasoning switched off, the model answers immediately and, xAI says, still performs strongly among non-reasoning systems.

How to use it

Grok 3 was being rolled out to users in the days after the announcement, with an early view of its reasoning. A Think button switches on the reasoning mode. In that mode the model exposes its working, so users can read the steps behind the answer rather than only the final result. No weight release or licence terms were described.

Questions people ask about Grok 3

What is Grok 3?

A language model from xAI, shown as a beta in February 2025. The company presents it as its most capable model, built for reasoning, maths, coding, world knowledge and instruction-following, with a smaller variant called Grok 3 mini.

How does the reasoning mode work?

Pressing Think turns on a chain-of-thought process shaped by large-scale reinforcement learning. The model can backtrack, simplify steps and check its own answer, spending anywhere from a few seconds to several minutes, and it shows the reasoning as well as the result.

What results does xAI report?

At its top test-time compute setting, Grok 3 (Think) reached 93.3% on the 2025 AIME, 84.6% on GPQA and 79.4% on LiveCodeBench. Grok 3 mini reached 95.8% on AIME 2024 and 80.4% on LiveCodeBench. An Elo of 1402 is reported in Chatbot Arena.

How do I get it?

xAI said it was rolling Grok 3 out to users in the days after the announcement, with an early preview of the reasoning features. No downloadable weights or licence terms were described.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Grok 3 yet, so the facts on this page come from the sources listed above: x.ai/news/grok-3. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.