Grok-2 and Grok-2 mini

Two models, one release

Grok-2 and Grok-2 mini are a pair of models released by xAI on 13 August 2024 in two sizes, offered at first as a beta for X Premium and Premium+ subscribers.^[1] The larger model, Grok-2, accepts images as input as well as text; xAI measured that ability with three vision benchmarks, MMMU, MathVista and DocVQA.^[1] The same release brought image generation to Grok on X, using Black Forest Labs' FLUX.1 model.^[1][2]

Splitting a release into a full model and a smaller sibling is a common pattern: the large model is aimed at the hardest tasks, while the smaller one is generally cheaper to serve and quick enough for everyday chat and coding help. Grok-2 mini is that smaller sibling in this pair.

Rollout

Grok-2 mini was available on the day of the announcement, while the full Grok-2 followed on 20 August 2024.^[2] The launch was a beta inside X: users reached the models through the Grok tab in the X app, and Grok-2 used real-time information from X.^[1] xAI said both models would also be offered through its enterprise API later in August 2024.^[1] That API, the company said, was built on a bespoke stack that allows multi-region inference, with mandatory multi-factor authentication, traffic statistics and billing analytics.^[1]

A beta label here means the models were live for subscribers but still being changed and fixed, rather than a finished general release; tying the first version to X's subscription tiers made the assistant part of the paid product.

Reported benchmark scores

xAI published a benchmark table for both models. For Grok-2 it reported GPQA 56.0%, MMLU 87.5%, MMLU-Pro 75.5%, MATH 76.1%, HumanEval 88.4%, MMMU 66.1%, MathVista 69.0% and DocVQA 93.6%.^[1] Grok-2 mini, the smaller model, scored GPQA 51.0%, MMLU 86.2%, MATH 73.0% and HumanEval 85.7%.^[1]

These benchmarks cover different skills: MMLU and MMLU-Pro are broad multiple-choice knowledge tests, GPQA is a set of hard graduate-level science questions, MATH is competition-style mathematics, HumanEval measures writing short programs, and MMMU, MathVista and DocVQA test reading images, charts and scanned documents. Scores like these are reported by the developer itself and describe behaviour on fixed test sets, not performance on every real task.

A model labelled "sus-column-r" was also tested anonymously on the LMSYS Chatbot Arena, where users compare two anonymous answers and vote. xAI said the model was Grok-2, and that it beat Claude 3.5 Sonnet and GPT-4 Turbo on overall Elo.^[1]

Updates and published weights

In December 2024 xAI shipped updated models, grok-2-1212 and grok-2-vision-1212, which it said had better accuracy, instruction following and multilingual ability.^[3] In August 2025 the company published the weights of Grok 2.5 on Hugging Face under the repository name xai-org/grok-2: about 500 GB spread over 42 files, needing 8 GPUs with more than 40 GB of memory each, released under the Grok 2 Community License Agreement, with commercial use restricted by an acceptable use policy.^[4][2]

Publishing weights lets other people download and run the model on their own hardware instead of calling an API, but the hardware requirements above show why few individuals do it: a full copy does not fit on a single consumer GPU. The licence matters too — downloadable weights are not the same as unrestricted use.

Where it sits in the line

xAI described Grok-2 as a significant step forward from Grok-1.5 in chat, coding and reasoning, and Grok-2 mini as a small but capable sibling.^[1] The successor, Grok 3, arrived in February 2025 and was trained with ten times the compute of Grok-2.^[2]