Hugging Face released gr.Workflow inside the Gradio library on August 25, 2026, introducing a drag-and-drop canvas designed to assemble, test, and deploy multi-step artificial intelligence pipelines. The feature converts interconnected Python steps into an interactive visual graph where intermediate outputs remain visible and individual nodes execute independently.
Every workflow graph relies on three node types: references for inputs, operators for computational steps, and subjects for outputs. Operators can run custom Python code, call external models through Hugging Face Inference Providers, route tasks to existing Gradio Spaces, or fetch records directly from dataset repositories on the Hub.
API generation and execution
Each subject node in a graph automatically generates a dedicated REST endpoint named after its label. Developers can query these endpoints via plain HTTP requests using curl or through the Python gradio_client library, passing authentication tokens for models hosted on Inference Providers. For workloads hosted directly inside a Space, functions decorated with @spaces.GPU run on ZeroGPU hardware, which allocates a GPU for the duration of the call and frees it upon completion.
Hugging Face published several reference Spaces demonstrating parallel and chained configurations. These examples include a single-node image editor using Qwen-Image-Edit, an art generator fanning out prompt requests to FLUX and large language models simultaneously, a parallel dataset analyzer built on the Datasets Server API, and an image animator running Lightricks/LTX-Video through Diffusers.
The company announced that an upcoming guide will demonstrate how to construct an AUTOMATIC1111-style image generation interface using the gr.Workflow system.
