| --- |
| 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](https://huggingface.co/spaces/EdgeFirst/Models) landing page. The visual interface is rendered from `index.html`. |
|
|
| Each model family lives in its own HuggingFace repo containing all size variants (nano through x-large) and platform-specific compiled formats. Models are trained and validated on [EdgeFirst Studio](https://edgefirst.studio), then published here. |
|
|
| --- |
|
|
| ## Model Repositories |
|
|
| ### Detection |
|
|
| | Repo | Model | Sizes | Nano mAP@0.5 | |
| |------|-------|-------|-------------| |
| | [EdgeFirst/yolo26-det](https://huggingface.co/EdgeFirst/yolo26-det) | YOLO26 | n/s/m | 55.0% | |
| | [EdgeFirst/yolo11-det](https://huggingface.co/EdgeFirst/yolo11-det) | YOLO11 | n/s/m | 53.1% | |
| | [EdgeFirst/yolov8-det](https://huggingface.co/EdgeFirst/yolov8-det) | YOLOv8 | n/s/m | 50.5% | |
| | [EdgeFirst/yolov5-det](https://huggingface.co/EdgeFirst/yolov5-det) | YOLOv5 | n/s/m | 47.8% | |
|
|
| ### Segmentation |
|
|
| | Repo | Model | Sizes | Nano Mask mAP | |
| |------|-------|-------|--------------| |
| | [EdgeFirst/yolo26-seg](https://huggingface.co/EdgeFirst/yolo26-seg) | YOLO26 | n/s/m | 32.7% | |
| | [EdgeFirst/yolo11-seg](https://huggingface.co/EdgeFirst/yolo11-seg) | YOLO11 | n/s | 30.2% | |
| | [EdgeFirst/yolov8-seg](https://huggingface.co/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 / YOLOv8 only; YOLO11 / YOLO26 not yet supported on eIQ Neutron)* |
| - **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](https://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 validation agent. Given a model and dataset, it runs full inference on the target device, captures per-image predictions in EdgeFirst Arrow/Parquet, and emits a Perfetto trace alongside. Loads each runtime 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 — so timing reflects deployed-application reality. |
| |
| ### EdgeFirst Validator |
| |
| The off-target post-processor. Consumes predictions + Perfetto trace, computes the 12-metric COCO accuracy tuple via `pycocotools` (or `lvis-api` for large-vocabulary datasets), and rebuilds per-stage timing summaries from the trace. Results attach to the Studio validation session as a structured YAML payload — the same payload this Model Zoo reads. |
| |
| ### EdgeFirst HAL |
| |
| The [EdgeFirst Hardware Abstraction Layer](https://github.com/EdgeFirstAI/hal) 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 |
| |
| Two timing surfaces per validation: |
| |
| - **`timing.inline`** — per-image `preprocess_ms` / `inference_ms` / `postprocess_ms` with min / mean / median / p95 / p99 / max. The universal contract every producer fills. |
| - **`timing.trace`** — full per-stage breakdown from the Perfetto trace (typically 25–33 stages), plus end-to-end FPS distribution. |
|
|
| Throughput exceeds the sum of stage latencies because the runtime pipelines I/O, preprocessing, NPU inference, and decode across frames. The Model Zoo headlines `trace.fps.median` as the throughput number, not the derived `1000 / (preprocess + inference + postprocess)`. Example: YOLOv5n on NXP i.MX 95 Neutron has per-stage means 21.7 + 12.2 + 15.8 ms (naive ≈ 20 FPS) but pipelined throughput of 56 FPS median. |
|
|
| --- |
|
|
| Validation results & card data: [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) · YOLO model weights: © Ultralytics Inc. (AGPL-3.0) · © 2026 [Au-Zone Technologies](https://www.au-zone.com) |
|
|
| <sub>NXP<sup>®</sup>, i.MX, eIQ<sup>®</sup>, 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.</sub> |
|
|
