README.md

metadata

title: EdgeFirst Model Zoo
emoji: 🔬
colorFrom: indigo
colorTo: red
sdk: static
pinned: true
license: cc-by-nc-4.0
short_description: Multi-platform model zoo validated on real edge hardware
thumbnail: https://huggingface.co/spaces/EdgeFirst/Models/resolve/main/social-card.png

EdgeFirst Model Zoo

This Space hosts the EdgeFirst Model Zoo landing page. The visual interface is rendered from index.html.

New here? Start with Introducing the EdgeFirst Model Zoo — what is in the zoo, why every figure links to its validation session, and how to reproduce any of it on your own hardware.

Want a platform, model or task we don't cover yet? Vote on what we measure next — three polls, one click each.

Each model family lives in its own Hugging Face repo containing all size variants (nano through x-large) and platform-specific compiled formats. Models are trained and validated on EdgeFirst Studio, then published here.

---

Model repositories

Detection

Repo	Model	Sizes	Nano mAP@0.5
EdgeFirst/yolo26-det	YOLO26	n/s/m	55.0%
EdgeFirst/yolo11-det	YOLO11	n/s/m	53.1%
EdgeFirst/yolov8-det	YOLOv8	n/s/m	50.5%
EdgeFirst/yolov5-det	YOLOv5	n/s/m	47.8%

Segmentation

Repo	Model	Sizes	Nano Mask mAP
EdgeFirst/yolo26-seg	YOLO26	n/s/m	32.7%
EdgeFirst/yolo11-seg	YOLO11	n/s	30.2%
EdgeFirst/yolov8-seg	YOLOv8	n/s/m	28.7%

---

Repo structure

Each model repo follows a consistent layout with platform folders:

EdgeFirst/yolov8-det/
├── README.md                                # Model card
├── onnx/
│   ├── yolov8n-det-fp32.onnx
│   └── ...
├── tflite/
│   ├── yolov8n-det-int8.tflite              # Default (logical split-decoder)
│   ├── yolov8n-det-int8-smart.tflite        # Smart variant
│   └── ...
├── imx95/
│   ├── yolov8n-det-int8.imx95.tflite
│   └── ...
├── hailo/
│   ├── yolov8n-det-int8.hailo8l.hef
│   └── ...
└── jetson/
    ├── yolov8n-det-fp16.orin-nano.engine
    └── ...

Naming convention

Pattern: {version}{size}-{task}-{precision}[-{variant}][.{platform}].{ext}

Component	Description	Examples
{version}{size}	Model family + variant	yolov8n, yolo11s, dfine-n
-{task}	Task suffix	-det, -seg, -semseg, -depth
-{precision}	Weight precision	-fp32, -fp16, -int8
-{variant}	Decoder variant (optional)	-smart
.{platform}	Deployment target (optional)	.imx95, .ara240, .hailo8l, .orin-nano
.{ext}	File format	.onnx, .tflite, .dvm, .hef, .engine

Decoder variants: No suffix = default for that format (logical split-decoder for INT8, combined for ONNX/float). -smart = multi-scale split-decoder offering better accuracy at higher compute cost.

Examples:

Description	Filename
ONNX FP32 (reference)	yolov8n-det-fp32.onnx
Generic INT8 TFLite	yolov8n-det-int8.tflite
Smart variant TFLite	yolov8n-det-int8-smart.tflite
NXP i.MX 95 TFLite	yolov8n-det-int8.imx95.tflite
Smart NXP i.MX 95	yolov8n-seg-int8-smart.imx95.tflite
Hailo-8L HEF	yolov8n-det-int8.hailo8l.hef
Jetson TensorRT FP16	yolov8n-det-fp16.orin-nano.engine

Supported hardware

• Linux x86_64 — ONNX Runtime CUDA / CPU (FP32 reference)
• Linux aarch64 — ONNX Runtime / TFLite (ARM64 generic Linux)
• Apple macOS — ONNX Runtime + CoreML ANE / GPU / CPU (FP16)
• NXP i.MX 8M Plus — 2.3 TOPS, TFLite INT8
• NXP i.MX 95 — 2.0 TOPS, eIQ Neutron TFLite (YOLOv5 and YOLOv8 published; YOLO11 and YOLO26 are in progress — they compile, load and run on the NPU while validation accuracy is resolved with NXP)
• NXP Ara240 — 40 eTOPS, .DVM
• RPi5 + Hailo-8L — 13 TOPS, HailoRT HEF
• NVIDIA Jetson Orin — 67–157 TOPS, TensorRT

Validation pipeline

Every artifact in the Model Zoo is measured on the same dataset on the same hardware users deploy on. Accuracy numbers and per-stage timing are produced by the same pipeline that runs the deployed model — there is no "benchmark configuration" separate from production.

End-to-end flow

Each training session produces a single set of weights in EdgeFirst Studio. The export pipeline emits ONNX FP32, INT8 TFLite, and platform-specific compiled formats (NXP i.MX 95 Neutron, NXP Ara240 .DVM, Hailo HEF, Jetson TensorRT). Every output is paired with an on-target validation that captures both accuracy (COCO mAP) and full-pipeline timing. The ONNX FP32 run from each training session is the reference baseline; quantization and runtime loss are measured relative to it.

EdgeFirst Profiler

The on-target measurement engine. Given a model and a dataset it runs the full pipeline on the target device — capture, preprocess, inference, postprocess — and computes the accuracy there too: COCO detection and segmentation metrics, the deployment confusion breakdown, and every timing block are calculated in-process, on the same machine that ran the model. There is no Python and no pycocotools on the device. It writes metrics.yaml, per-image predictions in EdgeFirst Arrow/Parquet, and a Perfetto trace.

Each runtime loads through its native delegate — VX Delegate on NXP i.MX 8M Plus, eIQ Neutron on NXP i.MX 95, NXP Ara SDK on Ara240, HailoRT on RPi5 + Hailo, TensorRT on Jetson, QNN on Qualcomm Hexagon, CoreML on Apple — so the timing reflects deployed-application reality rather than a benchmark harness.

EdgeFirst Studio

Where results are published, compared and browsed — not where they are computed. The profiler publishes its metrics, predictions, charts and trace to a v-XXXX validation session, and that session is what every figure in this Model Zoo cites. A run does not need Studio at all: point the profiler at a local model, a directory of images and a ground-truth file and it writes the same outputs to disk with nothing uploaded.

EdgeFirst HAL

The EdgeFirst Hardware Abstraction Layer provides hardware-accelerated primitives used at both validation and deployment time: letterbox resize, color-space conversion, normalization, layout conversion, YOLO/ModelPack post-decode, NMS. HAL automatically selects DMA-BUF, OpenGL ES, NXP G2D, or CPU paths depending on the platform. Apache 2.0; Rust + Python + C surfaces.

Latency and pipelined throughput

Per-frame latency is work-time — the sum of that frame's own stage durations — not wall-clock from capture to result. In an offline batch the capture workers race ahead of the bottleneck, so a wall-clock sojourn balloons with queue backlog and overstates per-image cost.

Throughput is measured separately, from the rate at which finished results actually emerge. The Model Zoo headlines realized FPS (realized_fps_scalar), the measured steady-state rate, not the derived 1000 / (preprocess + inference + postprocess). The two disagree because the runtime overlaps stages across frames.

Example: YOLOv5 Nano on NXP i.MX 95 eIQ Neutron, INT8, smart decoder, on an NXP FRDM-IMX95 board — per-stage means of 12.8 + 12.9 + 18.0 ms sum to 43.7 ms, which divides out to roughly 23 FPS. Measured throughput is 55.2 FPS (v-8ad), and the run is bound by postprocess rather than by the NPU.

---

Validation results & card data: CC BY-NC 4.0 · YOLO model weights: © Ultralytics Inc. (AGPL-3.0) · © 2026 Au-Zone Technologies

_{NXP^®, i.MX, eIQ^®, Neutron, and Ara240 are trademarks or products of NXP Semiconductors. Hailo is a trademark of Hailo Technologies Ltd. Jetson is a trademark of NVIDIA Corporation. All other trademarks are the property of their respective owners.}