NE301 Verified Models

NE301 is based on the STM32N6 NPU and supports TFLite Int8 quantized model inference. The tables below list all models verified on actual devices (12 models total).

Hardware Constraints

Parameter	Value	Parameter	Value
NPU	STM32N6, TFLite Int8	Max Resolution	480×480
Input Format	uint8, RGB888	Model Storage	10 MB (2 partitions, 5MB each)
Simultaneous Load	1 model	Output Quantization	`_uf`=float32, `_ui`=int8

Download file descriptions:

tflite — Original TFLite weight file for secondary development or model analysis

json — NE301 model configuration file defining input specs, post-processing type, and inference parameters

bin — NE301 OTA firmware package, deploy directly via the Web UI

1. General Object Detection (COCO 80 Classes)

Supports 80 categories including person, vehicle, animal, furniture, etc. _uf offers higher precision, _ui is ~10% faster.

Architecture	Size	Quant	Size	Verification Result	Model Files	OTA Package	License
YOLOv8n	256×256	_uf	3.2MB	● zebra 95% 1 detection	tflite json	bin	AGPL-3.0
YOLOv8n	256×256	_ui	3.2MB	● zebra 95% 1 detection	tflite json	bin	AGPL-3.0
ST YOLO-X	480×480	int8	2.4MB	● person 99.9% ● person 99.9% ● person 99.8% ● person 80% 4 detections	tflite json	bin	SLA0044

2. Person Detection

Detects only the "person" category with lighter post-processing.

Architecture	Size	Quant	Size	Verification Result	Model Files	OTA Package	License
YOLOv8n	256×256	_uf	3.1MB	● person 90% 1 detection	tflite json	bin	AGPL-3.0
YOLO11n	256×256	_uf	3.0MB	● person 92% 1 detection	tflite json	bin	AGPL-3.0

3. Pose Estimation

Each detected person outputs 17 keypoints (COCO keypoint format), suitable for action recognition and pose analysis scenarios.

Architecture	Size	Quant	Size	Verification Result	Model Files	OTA Package	License
YOLOv8n Pose	256×256	_uf	3.4MB	● pose 1 · 87% · 17 keys ● pose 2 · 81% · 17 keys ● pose 3 · 77% · 17 keys 3 poses	tflite json	bin	AGPL-3.0
YOLOv8n Pose	256×256	_ui	3.4MB	● pose 1 · 86% · 17 keys ● pose 2 · 81% · 17 keys ● pose 3 · 74% · 17 keys 3 poses	tflite json	bin	AGPL-3.0
YOLO11n Pose	256×256	_uf	3.6MB	● pose 1 · 80% · 17 keys ● pose 2 · 70% · 17 keys ● pose 3 · 63% · 17 keys 3 poses	tflite json	bin	AGPL-3.0

4. Meter Reading Detection

Supports two types of meter reading: digital meters and analog gauges. Digital meters recognize digits 0-9 on the display, suitable for water, electricity, and gas meter scenarios; the analog gauge model computes the pointer angle via 4 keypoints (dial center, max scale, min scale, pointer tip) and outputs the actual reading value, suitable for pressure gauges, thermometers, and other analog instruments.

Architecture	Size	Quant	Size	Verification Result	Model Files	OTA Package	License
YOLOv8n	256×256	_ui	3.1MB	● "0" 81% ● "0" 74% ● "0" 57% ● "0" 33% ● "3" 43% ● "5" 30% ● "5" 21% 7 detections	tflite json	bin	AGPL-3.0
YOLOv8n Pose	256×256	_ui	3.2MB	● gauge 92% · 4 keys ↻ reading 46.99 1 gauge · ratio 0.47 · ccw	tflite json	bin	AGPL-3.0

Analog gauge model note: The reading value is output via the Web UI and Webhook, and requires App ≥ v2.1.0.78 / Web UI ≥ v1.3.4.7 firmware.

5. Instance Segmentation (COCO 80 Classes)

Outputs pixel-level segmentation masks (32×32) for each detected object on top of object detection, suitable for fine-grained segmentation scenarios.

Architecture	Size	Quant	Size	Verification Result	Model Files	OTA Package	License
YOLOv8n Seg	256×256	_ui	3.6MB	● person 86% ● person 83% ● person 77% ● person 73% 4 segments	tflite json	bin	AGPL-3.0

6. Face Detection

Detects face positions and outputs 6 keypoints (left/right eye, nose tip, mouth center, left/right ear), suitable for face localization and expression analysis.

Architecture	Size	Quant	Size	Verification Result	Model Files	OTA Package	License
BlazeFace	128×128	_ui	245KB	● face 89% · 6 keypoints 1 face detected	tflite json	bin	Apache 2.0

Model Deployment

Two types of files are available for download from the tables above:

tflite + json: For custom development, requires self-compilation into a firmware package before deployment. See the Model Training and Deployment Guide
bin: Ready-to-use OTA firmware package, see the Quick Start Guide for details. Upload via Feature Debugging → upload or System Setting → Firmware Upgrade

More models are being adapted, stay tuned.

Document version: v1.7 · Last updated: 2026-06-12

Hardware Constraints​

1. General Object Detection (COCO 80 Classes)​

2. Person Detection​

3. Pose Estimation​

4. Meter Reading Detection​

5. Instance Segmentation (COCO 80 Classes)​

6. Face Detection​

Model Deployment​