xAI releases Grok-2 and Grok-2 mini in beta

Background

Before the release, an early version of the model had been tested in public under a codename, as described below.^[1]

What happened

On 13 August 2024 xAI released Grok-2 and Grok-2 mini in beta to X Premium and Premium+ subscribers. The company also promised an enterprise API later in the month.^[1] An early version had been tested on the LMSYS Chatbot Arena under the name "sus-column-r".^[1]

Key details

xAI reported Grok-2 at 87.5% on MMLU, 56.0% on GPQA and 88.4% on HumanEval.^[1] At the same time xAI began using Black Forest Labs' FLUX.1 model to expand Grok's capabilities on X.^[1] Grok used FLUX for image generation until xAI's Aurora model arrived in December 2024.^[2] Grok-2 itself reached users on 20 August 2024.^[2]

A beta release is a limited first step. It lets the developer see how a model behaves with real users before a wider launch. Betas are usually labelled as unfinished, so people know that answers and features may still change. Public arena tests add another kind of signal: anonymous users vote for the better of two answers, which gives a rough measure of how people judge a model against others. A vote-based rating is one of several ways to judge a model, and it does not replace formal benchmarks.

Readers can compare such results with the bare score of a test, but a vote tells only how one group of users preferred one answer on one day.

Impact

A year later, in August 2025, xAI published the weights of Grok 2.5 on Hugging Face under the Grok 2 Community License Agreement.^[3][2] Publishing weights makes a model available for others to download and run, which is different from an API that only the owner operates. The licence and its acceptable use policy still limit what users may do with the model.