Grok 2.5 (OSS Ver.)

The point of this release is access. xAI published the trained Grok 2 checkpoint as downloadable weights, so you can run it on your own machines instead of calling a hosted endpoint. The repository holds 42 files and about 500GB, and you load it locally with an inference engine rather than through a web page. For teams that need the model in-house, that is the whole value.

Grok 2.5 uses a mixture-of-experts design, where each layer holds multiple expert networks and only a couple activate per token. That keeps the compute per request lower than a dense model of the same total size, which matters when you are paying for GPUs by the hour. It is the same broad approach xAI used across the Grok line, and it is why the checkpoint can be large without being unusable.

The model accepts a 128K token context, enough to hold a long document in a single conversation. You can drop in a lengthy report, a code file or a transcript and ask questions across all of it without chunking first. For work like document review or long-form drafting, that window removes a lot of the plumbing you would otherwise build around a smaller model.

xAI points users to SGLang as the way to serve the model, and the repo's steps are specific. You run tensor parallelism across 8 GPUs, each with more than 40GB of memory, and apply FP8 quantization to fit the weights. The instructions assume a Linux server with that hardware, so the model is aimed at people who already run inference infrastructure rather than first-time users.

Because the weights are public, you can benchmark the model, inspect how its experts route and fine-tune it on your own data. That is different from a hosted chatbot, where you only get text in and text out. The release gives researchers a checkpoint from a named lab to test against, which is useful when published benchmark numbers are hard to reproduce otherwise.

Running Grok 2.5 yourself means prompts and outputs never leave your network. There is no external API call and no vendor seeing your traffic. For teams handling sensitive material, regulated data or internal code, that control is often the deciding factor between a hosted model and an open one, even when a hosted service would be cheaper.