Definition
Grok is a family of language and multimodal models introduced by xAI. The name also identifies the Grok assistant and related products. A Grok model produces outputs from supplied inputs; an assistant can add search, connectors, conversation state, and interfaces around that model.
An answer generated with X Search or Web Search uses information supplied by a tool. The API documentation requires those search tools to be enabled for current information.
Origin and attribution
The xAI team announced Grok in November 2023 and identified Grok-1 as the model powering the original assistant. The announcement credits the team.
On March 17, 2024, xAI released Grok-1's base weights and architecture under Apache 2.0. That checkpoint came from pretraining and had not been tuned for dialogue. Current API documentation is published under the SpaceXAI name.
Scope and limitations
Grok includes multiple checkpoints with their own capabilities and access terms. The Grok-1 release documents a mixture-of-experts base model; later releases need their own architecture and weight-access evidence.
The current API catalog distinguishes chat models from voice and media services. It also states that current information requires enabled search tools. A dated model identifier and a moving alias can resolve differently over time. Record which identifier, tools, and reasoning configuration produced a result.
Search access can retrieve mistaken or adversarial material, and a fluent answer can still contain false claims. The original announcement explicitly acknowledged false and contradictory output. Evaluate the exact deployment on the task it will perform.
Operational significance
Keep model selection separate from application permissions. A Grok-backed assistant may find an account or propose a change, while the factory still decides whether that action is allowed. An evaluation of the hosted assistant should preserve its tool configuration; testing the model without those tools measures a different system.
Distinguish it from nearby terms
- A chatbot is an application interface. Grok can name such an interface or the model behind it.
- An open-weight model distributes trained parameters. Grok-1's release supports that description for its checkpoint, without making it a family-wide access promise.
- A reasoning model describes inference behavior. Grok identifies a product family whose supported reasoning settings depend on the release.
Check your understanding
An evaluation says Grok answered today's news correctly. Which model identifier, search settings, retrieved sources, and application behavior must you know before reproducing that result?