Pre-trained, hardware-optimized, edge AI models for TI Microprocessor devices
The TI EdgeAI Model Hub is a curated repository of open-source computer vision models optimized for deployment on Texas Instruments Microprocessor devices.
Models are compiled for TI hardware using edgeai-tidl-tools or edgeai-tidlrunner, enabling production-ready inference without cloud dependency. For more details on TIDL model compilation options, runtimes, and supported operators, see the TIDL User Guide.
| Automotive | Aerospace & Defense | Industrial |
| Surveillance | Robotics | Edge IoT |
Models in this hub are distributed under various open-source licenses — each model's license is indicated in its own documentation page.
Disclaimer: Certain licenses in this repository impose distribution restrictions that may affect commercial, proprietary, or regulated-industry use. It is the sole responsibility of the user to review the applicable license terms, assess compatibility with their intended use, and obtain any necessary legal clearances prior to use or distribution. Texas Instruments makes no representation regarding the suitability of these licenses for any particular purpose and accepts no legal responsibility for the user's compliance obligations.
Compatible TI MPU device families compiled and validated via TIDL. See the supported devices, SDKs and version compatibility at the EdgeAI developer landing space and the edgeai-tidl-tools SDK version compatibility matrix.
| Device Family | Variants |
|---|---|
| AM62A | AM62A3 · AM62A7 |
| J722S | TDA4AEN · AM67A |
| J721E | TDA4VM |
| J721S2 | TDA4VE · TDA4VL · TDA4AL · AM68A |
| J784S4 | TDA4VH · TDA4AH · AM69A |
| Tool | Description |
|---|---|
| edgeai-tidlrunner | High-level compilation and benchmark interface. (Recommended for compilation and benchmark) |
| edgeai-tidl-tools | Deployment tools (and also low-level compilation tools for advanced users). |
1. Clone the repository
git clone https://github.com/TexasInstruments/edgeai-modelhub.git
cd edgeai-modelhub
2. Navigate to a model directory and prepare the model
cd models/vision/<task>/<model>/
python prepare_model.py --model <variant>
3. Compile for TI hardware (run from inside the edgeai-tidlrunner directory)
cd /path/to/edgeai-tidlrunner
tidlrunner-cli compile --target_device <device> \
--config_path /path/to/edgeai-modelhub/<model>_config.yaml
4. Infer on TI hardware (run from inside the edgeai-tidlrunner directory)
cd /path/to/edgeai-tidlrunner
tidlrunner-cli infer --target_device <device> \
--config_path /path/to/edgeai-modelhub/<model>_config.yaml
Deplyment can be done using edgeai-tidl-tools. For ONNX models, onnxruntime-tidl with TIDL acceleration can be used. Consult the documentation of edgeai-tidl-tools for more details.
| Model | Capability | Variants | Input | Reference Performance | License | Docs |
|---|---|---|---|---|---|---|
| MobileNetV3 | Image Classification | large | 224×224 | 75.3% Top-1 |  | View |
| ResNet-50 | Image Classification | v1.5, v1 | 224×224 | 74.93–76.15% Top-1 |  | View |
| DINO | Image Classification | ViT-S/16, ViT-S/8, ViT-B/16, ViT-B/8, ResNet-50 | 224×224 | 75.3–80.1% Top-1 | | View |
| DINOv2 | Image Classification | ViT-S/14, ViT-B/14 (w/ & w/o registers) | 224×224 | 80.9–84.6% Top-1 | | View |
| ViT | Image Classification | vit_b_16, vit_b_32, vit_l_16, vit_l_32 | 224×224 | 75.9–81.1% Top-1 | | View |
| ConvNeXt | Image Classification | convnext_tiny, convnext_small, convnext_base, convnext_large | 224×224 | 82.5–84.4% Top-1 | | View |
| DEIMv2 | Object Detection | s, m | 640×640 | 50.9–53.0% mAP | | View |
| DETR | Object Detection | detr_resnet50, detr_resnet50_dc5, detr_resnet101, detr_resnet101_dc5 | 800×800 (flexible) | AP50:95 42.0–44.9, AP50 62.4–64.7 | | View |
| Deformable-DETR | Object Detection | single-scale | 800×800 | AP50:95 39.4% | | View |
| RF-DETR | Object Detection | nano, s, m, l | 384–704px | 48.4–56.5% mAP | | View |
| RT-DETRv2 | Object Detection | s, ms, m, l, x | 640×640 | 48.1–54.3% mAP | | View |
| RTMDet | Object Detection | tiny, s, m, l, x | 640×640 | 40.9–52.8% mAP | | View |
| YOLO11 | Object Detection | n, s, m, l, x | 640×640 | 39.5–54.7% mAP |  | View |
| YOLO26 | Object Detection | n, s, m, l, x | 640×640 | 40.9–57.5% mAP | | View |
| YOLOv8 | Object Detection | n, m | 640×640 | 37.3–50.2% mAP | | View |
| YOLOX | Object Detection | nano, tiny, m, l, x, darknet53 | 416×416 / 640×640 | 24.8–51.2% mAP | | View |
Maintained by Texas Instruments EdgeAI Team | Last Updated August 2026
Pre-trained, hardware-optimized, edge AI models for TI Microprocessor devices
The TI EdgeAI Model Hub is a curated repository of open-source computer vision models optimized for deployment on Texas Instruments Microprocessor devices.
Models are compiled for TI hardware using edgeai-tidl-tools or edgeai-tidlrunner, enabling production-ready inference without cloud dependency. For more details on TIDL model compilation options, runtimes, and supported operators, see the TIDL User Guide.
| Automotive | Aerospace & Defense | Industrial |
| Surveillance | Robotics | Edge IoT |
Models in this hub are distributed under various open-source licenses — each model's license is indicated in its own documentation page.
Disclaimer: Certain licenses in this repository impose distribution restrictions that may affect commercial, proprietary, or regulated-industry use. It is the sole responsibility of the user to review the applicable license terms, assess compatibility with their intended use, and obtain any necessary legal clearances prior to use or distribution. Texas Instruments makes no representation regarding the suitability of these licenses for any particular purpose and accepts no legal responsibility for the user's compliance obligations.
Compatible TI MPU device families compiled and validated via TIDL. See the supported devices, SDKs and version compatibility at the EdgeAI developer landing space and the edgeai-tidl-tools SDK version compatibility matrix.
| Device Family | Variants |
|---|---|
| AM62A | AM62A3 · AM62A7 |
| J722S | TDA4AEN · AM67A |
| J721E | TDA4VM |
| J721S2 | TDA4VE · TDA4VL · TDA4AL · AM68A |
| J784S4 | TDA4VH · TDA4AH · AM69A |
| Tool | Description |
|---|---|
| edgeai-tidlrunner | High-level compilation and benchmark interface. (Recommended for compilation and benchmark) |
| edgeai-tidl-tools | Deployment tools (and also low-level compilation tools for advanced users). |
1. Clone the repository
git clone https://github.com/TexasInstruments/edgeai-modelhub.git
cd edgeai-modelhub
2. Navigate to a model directory and prepare the model
cd models/vision/<task>/<model>/
python prepare_model.py --model <variant>
3. Compile for TI hardware (run from inside the edgeai-tidlrunner directory)
cd /path/to/edgeai-tidlrunner
tidlrunner-cli compile --target_device <device> \
--config_path /path/to/edgeai-modelhub/<model>_config.yaml
4. Infer on TI hardware (run from inside the edgeai-tidlrunner directory)
cd /path/to/edgeai-tidlrunner
tidlrunner-cli infer --target_device <device> \
--config_path /path/to/edgeai-modelhub/<model>_config.yaml
Deplyment can be done using edgeai-tidl-tools. For ONNX models, onnxruntime-tidl with TIDL acceleration can be used. Consult the documentation of edgeai-tidl-tools for more details.
| Model | Capability | Variants | Input | Reference Performance | License | Docs |
|---|---|---|---|---|---|---|
| MobileNetV3 | Image Classification | large | 224×224 | 75.3% Top-1 | | View |
| ResNet-50 | Image Classification | v1.5, v1 | 224×224 | 74.93–76.15% Top-1 | | View |
| DINO | Image Classification | ViT-S/16, ViT-S/8, ViT-B/16, ViT-B/8, ResNet-50 | 224×224 | 75.3–80.1% Top-1 | | View |
| DINOv2 | Image Classification | ViT-S/14, ViT-B/14 (w/ & w/o registers) | 224×224 | 80.9–84.6% Top-1 | | View |
| ViT | Image Classification | vit_b_16, vit_b_32, vit_l_16, vit_l_32 | 224×224 | 75.9–81.1% Top-1 | | View |
| ConvNeXt | Image Classification | convnext_tiny, convnext_small, convnext_base, convnext_large | 224×224 | 82.5–84.4% Top-1 | | View |
| DEIMv2 | Object Detection | s, m | 640×640 | 50.9–53.0% mAP | | View |
| DETR | Object Detection | detr_resnet50, detr_resnet50_dc5, detr_resnet101, detr_resnet101_dc5 | 800×800 (flexible) | AP50:95 42.0–44.9, AP50 62.4–64.7 | | View |
| Deformable-DETR | Object Detection | single-scale | 800×800 | AP50:95 39.4% | | View |
| RF-DETR | Object Detection | nano, s, m, l | 384–704px | 48.4–56.5% mAP | | View |
| RT-DETRv2 | Object Detection | s, ms, m, l, x | 640×640 | 48.1–54.3% mAP | | View |
| RTMDet | Object Detection | tiny, s, m, l, x | 640×640 | 40.9–52.8% mAP | | View |
| YOLO11 | Object Detection | n, s, m, l, x | 640×640 | 39.5–54.7% mAP | | View |
| YOLO26 | Object Detection | n, s, m, l, x | 640×640 | 40.9–57.5% mAP | | View |
| YOLOv8 | Object Detection | n, m | 640×640 | 37.3–50.2% mAP | | View |
| YOLOX | Object Detection | nano, tiny, m, l, x, darknet53 | 416×416 / 640×640 | 24.8–51.2% mAP | | View |
Maintained by Texas Instruments EdgeAI Team | Last Updated August 2026