NVIDIA Releases TensorRT Model Connect in Public Preview
NVIDIA has released TensorRT Model Connect (TRTMC) in public preview as an Apache-2.0 open-source project. It converts supported Hugging Face or local checkpoints into TensorRT inference through two commands, without an intermediate ONNX export. The build creates a versioned .bundle file that can be loaded by native C++ task APIs, allowing inference in services, embedded applications and robotics systems without PyTorch in the runtime path.
TRTMC uses model-family-specific reference implementations rather than one universal converter. Python handles checkpoint resolution and TensorRT engine construction, while C++ executes inference from the bundle. Available task APIs include generate(), transcribe(), generate_image(), embed() and solve(). The trtmc inspect command shows the bundle type, model family, precision, runtime identity and engines. Some hybrid profiles require a helper Python executable, with that dependency recorded in their manifests.
Current release wheels support Linux aarch64, Python 3.10 or 3.12, glibc 2.39 or newer, and TensorRT 11.1.0.106. No x86_64 wheels are available; those users must build from source in Docker. A July 29, 2026 GB300 snapshot lists 105 profiles across 76 families, with 102 exceeding their declared reference by more than 5%, according to NVIDIA.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.