Model zoo and compiler front end for Hailo’s edge AI silicon

What it does The Hailo Model Zoo bundles pre-trained computer vision models—classification, object detection, and segmentation—with a full build and evaluation environment for Hailo’s edge AI accelerators. It handles parsing ONNX or TensorFlow models, optimizing and quantizing them, compiling to Hailo Executable Format binaries, and benchmarking accuracy on both an emulator and physical hardware. It also ships retraining instructions for adapting models to custom datasets.

The interesting bit Rather than dumping checkpoints, the repository acts as the front end to Hailo’s proprietary toolchain, abstracting the NPU-specific quantization and compilation workflow. The trade-off is tight vendor lock-in: the core Dataflow Compiler and runtime are closed source and require a customer relationship with Hailo to access.

Key highlights

• Public models and in-house “Hailo Models” spanning Hailo-8, Hailo-8L, Hailo-10H, Hailo-15H, and Hailo-15L hardware targets
• End-to-end pipeline from model parsing through compilation to HEF generation, plus emulator and on-device accuracy evaluation
• Recent additions include YOLOv11-obb, NMS-free YOLO26, PaddleOCR-v5, and YOLOv5-seg with HailoRT postprocessing
• MIT-licensed code, though issue reporting is disabled and redirected to the Hailo Community Forum

Caveats

• The master branch supports only Hailo-10 and Hailo-15 devices; Hailo-8 and Hailo-8L require the legacy v2.x branch and an older compiler stack
• Full functionality depends on the proprietary Hailo Dataflow Compiler and optional HailoRT, which are not included and require contacting Hailo for access
• Issue reporting is disabled on the repository; bugs and feature requests must go through the external Hailo Community Forum

Verdict Useful if you are already committed to Hailo silicon and need validated, quantize-ready vision models. Skip it if you want vendor-agnostic tooling or a fully open-source edge deployment stack.