Hailo Model Zoo: pre-trained models and the HEF compilation path for Hailo-10 and Hailo-15
The Hailo Model Zoo includes pre-trained models and a full building and evaluation environment
At a glance
- What is it?
- The Hailo Model Zoo is a Python toolchain that wraps the Hailo Dataflow Compiler to parse, optimize, compile and evaluate quantized vision models for Hailo accelerators. Its main constraint is not code but licensing: the compiler and runtime are commercial products you must obtain from Hailo before most commands do anything.
- Who is it for?
- Adopt it if you already hold a Hailo Dataflow Compiler licence and are targeting Hailo-10 or Hailo-15 on the master branch, or Hailo-8 and Hailo-8L on the v2.x branch. Do not adopt it if you have no Hailo hardware or no compiler licence, because the README routes non-customers to a contact form before installation begins, and a model zoo that cannot compile is just a folder of ONNX files.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 13 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The gap between a trained network and a Hailo accelerator
A PyTorch or TensorFlow checkpoint does not run on a Hailo device. Something has to translate the graph into the compiler's internal representation, choose quantization parameters, and emit a binary the device can execute. The Hailo Model Zoo is that something, packaged with a catalogue of models that have already been through the process. The README describes four stages: parse, optimize, compile, evaluate. Parse converts the input model into Hailo's internal representation. Optimize performs inference optimization and produces a compressed integer representation. Compile invokes the Hailo compiler to produce the Hailo Executable Format (HEF) file. Evaluate runs the model on the Hailo Emulator or on real hardware and reports accuracy. The audience is embedded and edge engineers who have picked a Hailo part and now need a working HEF plus a defensible accuracy number. It is not a training framework and it is not a general ONNX converter; it is the vendor's own path from checkpoint to silicon.
Two model families and a branch split that decides your silicon
The catalogue splits into public models trained on publicly available datasets and Hailo Models trained in-house on internal datasets for specific use cases. Public models are documented per device: Hailo-15H, Hailo-10H and Hailo-15L each get their own classification, object detection and semantic segmentation pages, with a shared page for other tasks. The compatibility notice in the README is the part to read twice. Hailo-8 and Hailo-8L are supported on the Hailo Model Zoo v2.x branch combined with the Hailo Dataflow Compiler v3.x branch, and the master branch is intended for Hailo-10 and Hailo-15 devices only. That is a hard fork in practice, not a documentation footnote. If you clone master and try to compile for a Hailo-8, the model pages you need are on a different branch and a different compiler generation. The release list reflects this: v5.4.0 and v5.3.0 sit alongside a v2.19.0 published on the same day as v5.4.0, which is consistent with two maintained lines rather than one. Each Hailo Model also ships retraining instructions for custom datasets, covered in docs/RETRAIN_ON_CUSTOM_DATASET.rst.
Installation starts with a licence, not a pip command
The Quick Start Guide lists its prerequisites in order, and the first two are not open source. You install the Hailo Dataflow Compiler and enter the virtualenv, and the README states that if you are not a Hailo customer you should contact hailo.ai. HailoRT is optional and required only to run on Hailo hardware, again gated behind the same contact route. Only then does the repository itself install: clone https://github.com/hailo-ai/hailo_model_zoo.git, then cd hailo_model_zoo and pip install -e . . The README's own smoke test is hailomz info mobilenet_v1, which prints model information without touching hardware. The environment the badges advertise is Python 3.10, 3.11 or 3.12, TensorFlow 2.19.1, CUDA 12.5.1, Dataflow Compiler 5.4.0 and HailoRT 5.4.0. That is a narrow, opinionated stack. TensorFlow 2.19.1 and CUDA 12.5.1 are specific enough that a host already running a different CUDA generation will need a separate environment rather than an upgrade. The README also mentions three Claude Code skills that walk through the full DFC flow on a fresh model, so agent-assisted workflows are an intended entry point rather than an afterthought.
What the emulator can and cannot tell you
The README distinguishes three accuracy measurements: full precision accuracy of each model, quantized accuracy using the Hailo Emulator, and accuracy measured on Hailo's device. The first two run without hardware. The third requires HailoRT and a physical part. That ordering matters when you plan a project, because the gap between quantized emulation and device results is the number that decides whether a model is deployable, and the README does not claim the emulator closes it. It presents emulator accuracy as a distinct figure from device accuracy, which is the honest framing. The profiler stage produces a report containing information about the model and expected performance on the Hailo hardware. Note the word expected. The README does not present profiler output as a guarantee, and anyone treating it as a benchmark should reproduce measurements using the process in docs/BENCHMARKS.rst, which the README points to rather than inlining.
Where this is the wrong tool
If your target is not Hailo silicon, nothing here applies. The compile stage emits HEF, a format only Hailo devices execute, and the evaluation stages depend on the Hailo Emulator or HailoRT. There is no path in the described flow that produces a portable artifact. The second failure mode is version drift. The release notes for v5.4.0 describe the update as aligning with the latest Hailo software stack, upgrading to Dataflow Compiler v5.4.0 and HailoRT v5.4.0. The model zoo therefore moves in lockstep with a commercial stack you do not control, and a compiler upgrade can invalidate previously generated HEF files or change quantization behaviour. Teams that pin an older model zoo against a newer compiler, or the reverse, are outside the configuration the README describes. Third, the licence situation is asymmetric: the repository is MIT, but the Dataflow Compiler and HailoRT are separate products obtained through Hailo. The MIT grant covers the model zoo code, not the toolchain it drives, and not the weights or dataset terms attached to individual models. That distinction deserves a read of each model page rather than an assumption from the repository licence badge.
How this differs from a general ONNX to edge compiler
Apache TVM and the OpenVINO toolkit both take a trained model and target edge hardware, but they are organized around a hardware-agnostic compiler with vendor backends contributed to it. The Hailo Model Zoo is organized the other way round: a fixed catalogue of models, each with a documented accuracy figure, wrapped around one vendor's compiler. The practical difference shows up in what you get out of the box. With a general compiler you bring your own architecture and spend your time on operator coverage and scheduling. With the model zoo you start from a model that has already been parsed, optimized and compiled for the target device, and the README's retraining instructions let you adapt it to a custom dataset instead of porting an arbitrary graph. That is a narrower offer, and it is also a shorter path when one of the catalogued architectures fits the task. If your model is genuinely novel, the catalogue is a constraint rather than a convenience.
Upgrade cost and what the MIT licence does not cover
The maintenance story is version pairing. Each model zoo release names the Dataflow Compiler and HailoRT versions it aligns with, and v5.4.0 names 5.4.0 for both. Upgrading the compiler means upgrading the model zoo, re-running the compile stage, and re-checking quantized accuracy, because the integer representation produced by the optimize stage is compiler output, not a stable file format you can carry forward. The v2.19.0 release appearing on the same date as v5.4.0 indicates that the older Hailo-8 line continues to receive releases, so teams on Hailo-8 are not stranded, but they are on a parallel branch with its own compiler generation. On licensing, the repository is MIT, and the README links the LICENSE file for that. The Dataflow Compiler and HailoRT are not part of that grant; the README directs non-customers to Hailo's contact page for both. Model weights and the datasets behind them may carry their own terms, and the README does not enumerate those per model. This is a description of what the material states, not legal advice; if you are shipping a product, have counsel review the compiler agreement and each model's page.
Editorial conclusion
Adopt it if you already hold a Hailo Dataflow Compiler licence and are targeting Hailo-10 or Hailo-15 on the master branch, or Hailo-8 and Hailo-8L on the v2.x branch. Do not adopt it if you have no Hailo hardware or no compiler licence, because the README routes non-customers to a contact form before installation begins, and a model zoo that cannot compile is just a folder of ONNX files. Before writing any integration code, verify two things: which branch matches your silicon, since the README states the master branch is intended for Hailo-10 and Hailo-15 only, and that the Dataflow Compiler and HailoRT versions you install match the v5.4.0 pair named in the release notes, because the model zoo tracks the stack version rather than pinning an independent one.
Community notes