xAI open sources the Grok-1 base model, training code stays closed

xAI has open sourced the base model behind Grok (the 314-billion-parameter Mixture-of-Experts model the company says sits under the chatbot), with weights published on GitHub under the Apache License 2.0. One part is deliberately missing: there is no training code, and xAI describes the training stack as "custom" without giving details.

What actually shipped

In a blog post, xAI says the release covers the model's base weights, not a ready-made assistant. The company notes the model "wasn't tuned for any particular application," conversation included. In practice, what you get is a base model: the fine-tuning or instruction work that makes it behave like the Grok you may have seen is left to you.

The license matters if you're planning a product. Apache 2.0 permits commercial use, so this is usable in shipped software, not only in research, the same license family many other open-weights models circulate under. The post publishes no benchmark numbers. TechCrunch's reporting carries the 314B parameter count, but not how many parameters are active per token, usually the number that sets serving cost for a mixture-of-experts model, and the one to ask for before budgeting compute.

What's missing

Three things are not in the release. First, the training code and pipeline details: xAI calls the stack "custom" and stops there. Second, any tie-in with X: the Grok chatbot, which went to Premium+ subscribers last year, can reach some data from the X network; the open model has no connection to the social network. Third, the application-layer tuning that turns the base model into a chatbot in the first place.

That split (weights out the door, training pipeline and data connections kept in-house) is a narrower move than a full open release, and a wider one than the API-only access most frontier labs offer.

Where Grok-1 lands among the open models

The release puts xAI in the company of a growing list of labs that have published open-weights models: Meta's LLaMA, Mistral, Falcon, AI2's models, and, in February, Google's Gemma 2B and Gemma 7B. The practical difference for you is the usual open-weights math: if you want to self-host, fine-tune, or build on a specific model's behavior, the weights being available is the precondition, and Grok-1 now clears it under a permissive license.

At least one team already has a plan. Perplexity CEO Aravind Srinivas posted on X that the company will fine-tune Grok for conversational search, optimize the inference, and make it available to Perplexity's Pro users. Whether that holds up under a license review is Perplexity's problem, not xAI's.

Why now

Musk telegraphed the move: last week he noted on X that xAI intended to open source the Grok model that week. The timing also lands in the middle of Musk's legal fight with OpenAI; he sued the company earlier this month over what he called a betrayal of its nonprofit founding goals, and has since called out OpenAI and Sam Altman publicly several times. The lawsuit is a story of its own; it doesn't change what xAI shipped, but it colors how the release is being read.

What it means if you're choosing a base model

If you're evaluating base models to fine-tune or serve, Grok-1 adds an Apache 2.0 option at the top of the self-hosting scale range. Without published benchmarks or a reference training recipe, the honest path is to benchmark it against the alternatives you're already considering before committing compute. The weights are the entry ticket; the missing training code means xAI is publishing that you can build on the model, not how the model was made.