Roboflow docs use some terms, and some terms of art, that may be unfamiliar. This page clarifies their meaning in the context of the Roboflow platform and Roboflow Inference.
API
An interface that allows software applications to communicate with each other. The Roboflow API provides access to datasets, models, Workflows, and inference, so you can configure and execute pipelines programmatically. See the REST API reference.
Block
The fundamental unit of a Workflow. Blocks perform specific tasks, such as running model inference, performing logic, or interfacing with external services.
CLI
Command-line interface. See the Roboflow CLI for platform operations and the Inference CLI for starting a server, running benchmarks, and executing Workflows.
Client
A software application or library that interacts with an API. The Inference SDK is a client that abstracts the REST API into a friendlier interface.
Commercial license
A license permitting businesses to use models under terms suited for commercial applications, typically involving subscription plans or usage fees. Inference itself uses a permissive Apache 2.0 license, but some models have terms that may require downstream users to open source their own codebase; a commercial license removes that restriction.
Dedicated
A configuration where cloud compute resources are allocated exclusively to a single user or organization, ensuring consistent performance and enabling additional functionality. Billed based on running time. See Dedicated Deployments.
Definition
The structured JSON description of a Workflow, including its sequence of Blocks, input parameters, response format, and any custom logic.
Dynamic Block
A Workflows Block using custom Python code included in the JSON definition of the Workflow, allowing advanced customization at runtime.
Enterprise
A service tier for large organizations, offering enhanced capabilities, scalability, and support. Source code for enterprise Inference functionality lives in the enterprise folder of the repo but may only be used with an active Enterprise license. See Enterprise Deployment.
Execution Engine
The backend system responsible for executing Workflows, managing the execution of Blocks, and optimizing resource allocation.
Fine-tuned model
A model trained on a specific dataset for improved performance in a targeted application. For example, a "scratch detection" model tuned to find scratches on a specific automotive component.
Foundation model
A general-purpose model that knows a lot about a lot, does not necessarily need to be fine-tuned on a specific dataset, and can be used zero-shot. See Supported Models.
Kind
A categorization of data types used in Workflows. Defining the input and output Kinds of a Workflow Block allows the Execution Engine to validate, serialize, and optimize data and connections. Example Kinds are detection (a prediction from an object detection model), image (pixels and their metadata), and float_zero_to_one.
LMM
Large multimodal model: a model that processes multiple data types, such as images and text. Examples are Florence-2 and GPT-4o.
Managed
A deployment running in Roboflow's cloud environment, where scaling and infrastructure are provided as a service. Contrast with self-hosted, where you install the software on your own infrastructure and are responsible for its setup and maintenance.
Metered
A billing approach based on usage metrics, such as the number of Workflow executions or hours of video processed. Models and Workflows that require an API key to access Roboflow's cloud services are metered.
Model
A trained machine learning artifact used for inference tasks such as object detection or classification. Consists of an architecture (like ResNet-32) and trained weights.
Parameter
An input to a Workflow. Can be an image or data such as strings, numbers, arrays, or objects. Used as inputs to Blocks.
Platform
The Roboflow ecosystem, which includes end-to-end tools for collecting and organizing data, annotating images, creating datasets, training models, building Workflows, and monitoring deployments.
Pre-trained model
A model architecture loaded with weights trained on a generic dataset (like Microsoft COCO). Contrast with a fine-tuned model, which has been trained on a domain-specific dataset.
Public
A dataset, model, or Workflow accessible to all users within the Roboflow ecosystem. These can be found and distributed on Roboflow Universe.
Schema
The structured format of input and output data for a Workflow, defining the properties and types a downstream application needs to pass and parse.
SDK
Software development kit: a set of tools and libraries for integrating with an API using user-friendly abstractions. See the Python SDK and the Inference SDK.
Server
A device running the Inference HTTP interface, usually through Docker on port 9001. See Self-Hosted Deployment.
Serverless
An execution model where resources scale on demand, without manual resource management. Contrast with dedicated. Billed solely based on usage. See the Serverless Hosted API.
Traditional CV
Blocks in Workflows that implement traditional computer vision techniques, such as filtering or edge detection, without relying on machine learning models.
Universe
A large collection of datasets, pre-trained models, and other resources shared publicly by other users and available within the Roboflow ecosystem. See Roboflow Universe.
Weights
The learned parameters of a machine learning model, which determine its predictions for a given input.
Workflow
A series of interconnected Blocks designed to process data and produce outcomes through a defined sequence of operations. Used for chaining models, maintaining state, performing custom logic, and integrating with external systems. See Workflows.
Workspace
A collaborative environment in Roboflow where users create datasets, train models, and build Workflows. A Workspace is the container for data and the arbiter of access via seats and API keys.