Grok 2 — open-weight release of xAI's 2024 production model

TL;DR

xAI released the weights of Grok 2 — the production model used at xAI in 2024 — under a custom “Grok 2 Community License” on Hugging Face. The checkpoint is ~500 GB across 42 files, post-trained, and shipped as a TP=8 sharded artifact intended to be served at FP8 on 8 GPUs (≥40 GB each) via SGLang ≥ v0.5.1 with the Triton attention backend. No architecture details, no training data disclosure, no tech report, no benchmarks — the model card is purely a “how to serve it” page. Notable as a delayed open-weight release of a previous-generation frontier closed model, in the same lineage as Meta’s earlier Llama releases and OpenAI’s gpt-oss.

Key claims

Weights are licensed under a custom “Grok 2 Community License Agreement” rather than a standard open-source license such as Apache 2.0 or MIT [§License].
The checkpoint is approximately 500 GB across 42 files, downloaded via hf download xai-org/grok-2 [§Usage].
The released artifact is pre-sharded for TP=8 inference — the user is expected to run it on 8 GPUs, each with >40 GB of memory, and the documented serving recipe uses --quantization fp8 --attention-backend triton under SGLang ≥ v0.5.1 [§Usage].
The checkpoint is post-trained; a specific chat template (mirroring the SGLang tiktoken tokenizer’s Human: … <|separator|>\n\nAssistant: format) is required to elicit correct behavior [§Usage].
The model card contains no architecture, training data, evaluation, or safety information — only a serving recipe and a license link [§Grok 2, §License].

Method

xAI did not document the model. The page describes only deployment: download the 500 GB / 42-file checkpoint, install SGLang ≥ v0.5.1, launch a TP=8 FP8 inference server with the Triton attention backend, and send a request with the post-trained chat template. The choice of TP=8 + FP8 + Triton attention is consistent with a large dense or MoE checkpoint sized for ~8× H100-class GPUs at inference, but the model card itself does not confirm which.

Results

None reported on the model card. There are no benchmark numbers, ablations, or comparisons to other models. The only quantitative metric on the page is the Hugging Face “Downloads last month” counter (45,315 at the time of fetching) [§Downloads last month].

Why it’s interesting

A delayed-release of a previous-generation production frontier model is a distinct release shape from what this wiki has been tracking — most open releases here (e.g. Kimi K2: Open Agentic Intelligence, GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models, HunyuanImage 3.0 Technical Report, DeepSeek-V3.1-Base — long-context-extended V3 base with UE8M0 FP8 microscaling) ship at the current frontier with a tech report and benchmarks. Grok 2 ships only weights + a serving recipe under a custom license, with no documentation: it’s a useful baseline / capability-snapshot artifact rather than a deployment-ready open release in the Open foundation-model releases sense. Worth tracking whether other labs follow this pattern (release N-1 weights when N is the production model) and whether the custom-license precedent shapes future releases.