Grok-2
xAI's 2024 chat and reasoning model, with a smaller sibling, Grok-2 mini
Grok-2 is a large language model from xAI, released in 2024 as the successor to Grok-1.5, with a smaller sibling called Grok-2 mini. It handles chat, coding, reasoning and vision tasks, and was first offered inside the X app before its weights were published under a community licence.
- Succeeded Grok-1.5 in 2024, with a smaller sibling, Grok-2 mini
- Tested anonymously on Chatbot Arena as "sus-column-r"
- Weights are about 500 GB across 42 files, under a community licence
- Serving setup needs 8 GPUs with over 40GB memory each
- Handles text and vision, with real-time information from X
What it is
Grok-2 is a language model built by xAI, announced in 2024 as a step beyond the earlier Grok-1.5. It is aimed at chat, coding and reasoning work, and a smaller companion model, Grok-2 mini, trades some quality for speed. The company describes Grok-2 as its assistant with both text and vision understanding, wired to live information from the X platform. Trained and used at xAI during 2024, its weights were later published on Hugging Face as a post-trained checkpoint of roughly 500 GB, licensed under the Grok 2 Community License Agreement.
How it came about
xAI introduced Grok-2 and Grok-2 mini as an early preview, following Grok-1.5 and the first Grok model announced in November 2023. Before the announcement, an early build was entered into the LMArena.ai Chatbot Arena under the code name "sus-column-r". The models began as a beta on X for Premium and Premium+ subscribers, with an enterprise API promised for later that month. xAI said its small team was moving quickly and focusing on core reasoning with a new compute cluster.
Inside the model
The published checkpoint is a post-trained model, meaning it was further tuned after pre-training rather than left raw. Serving instructions show it is set up for tensor parallelism across eight GPUs, each needing more than 40GB of memory, with fp8 quantisation and a tiktoken-style tokenizer file. The download spans 42 files and about 500 GB. Internally, xAI says it judges outputs using AI Tutors who compare two candidate responses against written guidelines, focusing on instruction following and factual accuracy. No parameter count, training data or compute figures are given.
What it can do
xAI reports gains over Grok-1.5 on academic tests covering reasoning, reading, maths, science and coding, with results it calls competitive on GPQA, MMLU, MMLU-Pro and MATH. The company also claims strong visual maths reasoning on MathVista and document question answering on DocVQA. On Chatbot Arena, the "sus-column-r" build reportedly beat Claude 3.5 Sonnet and GPT-4-Turbo on overall Elo. Improvements are noted in reasoning over retrieved content and tool use, such as spotting missing information and ignoring irrelevant posts.
How to use it
Grok-2 and Grok-2 mini reached users through the Grok tab in the X app for Premium and Premium+ subscribers, and were planned for an enterprise API with multi-region inference, multi-factor authentication and billing analytics. The weights are downloadable from Hugging Face under the Grok 2 Community License Agreement, and can be served with the SGLang engine using the provided chat template.
Safety and testing
xAI describes an internal evaluation process in which AI Tutors compare paired model responses against written criteria, concentrating on following instructions and giving accurate, factual information. The company also highlights security features on its enterprise API, including mandatory multi-factor authentication. No separate safety evaluations, red-teaming details or incidents are described.
Reception and debate
An early version tested anonymously on the LMArena.ai Chatbot Arena drew attention by reaching a higher overall Elo score than Claude 3.5 Sonnet and GPT-4-Turbo at the time of the announcement, according to xAI. The company presented the model as placing it among the leaders in AI development. No independent reviews, criticism, lawsuits or bans are described here.
Why it matters
Grok-2 continued the Grok line that began with Grok-1 in November 2023 and pushed xAI into vision as well as text. xAI framed it as groundwork for AI-driven features on X, such as better search, deeper reading of posts and improved replies, and said a preview of multimodal understanding would follow for both the X experience and the API. The release of the weights under a community licence also made the model available to developers outside the platform.
Questions people ask about Grok-2
What is Grok-2?
It is a language model from xAI, released in 2024 as the follow-up to Grok-1.5. It covers chat, coding and reasoning, and can also work with images. A smaller version, Grok-2 mini, is built for faster answers with somewhat lower quality.
How can I use it?
Subscribers to X Premium and Premium+ could reach it through the Grok tab in the X app during the beta. xAI also planned an enterprise API, and the weights are downloadable from Hugging Face under the Grok 2 Community License Agreement for self-hosting.
How big is the model?
No parameter count is given. What is published is the storage footprint: the download contains 42 files and comes to roughly 500 GB. Serving the checkpoint as provided requires eight GPUs, each with more than 40GB of memory, using tensor parallelism.
How well does it perform?
xAI reports improvements over Grok-1.5 across reasoning, reading, maths, science and coding tests, and claims strong results on GPQA, MMLU, MMLU-Pro, MATH, MathVista and DocVQA. An early build also reached a higher Elo than Claude 3.5 Sonnet and GPT-4-Turbo on Chatbot Arena at the time.
What can it see?
Grok-2 understands images as well as text. xAI highlights visual maths reasoning and answering questions about documents, and it experimented with Black Forest Labs' FLUX.1 model to add further capabilities on X.
Good reading and sources
There is no English Wikipedia article on Grok-2 yet, so the facts on this page come from the sources listed above: x.ai/news/grok-2, huggingface.co/xai-org/grok-2. 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.