TensorRT-Model-Connect
A public-preview toolkit that builds a Hugging Face model into a versioned TensorRT inference bundle in two commands, with C++ task APIs for text generation, transcription, image and video generation, segmentation, embedding and forecasting.
TensorRT-Model-Connect takes a supported Hugging Face or local checkpoint and builds TensorRT engines directly, without an intermediate ONNX export step, producing a versioned bundle file. That bundle can then be run from native C++ task APIs, so the same build works both from a Python-first build environment and from a C++ application.
It is described as an extensive collection of AI model reference implementations in C++ on top of NVIDIA TensorRT, kept current by an agentic workflow that continually adds support for new models. The project positions itself for quickly exploring models and evaluating broad model coverage, and points to a separate project, TensorRT Edge-LLM, for production deployment on NVIDIA edge platforms where performance is the priority. It is in public preview, so its APIs, scope and direction may still change.