Verified asset on Archimedes Market. View the full 4-dimension Trust Report (security · quality · license · complexity) and the curated catalog on the asset page.

---

Lightweight Edge ML Model Zoo

Six pre-quantized inference models sized for <256 KB RAM and <1 MB flash, with measured latency on three reference SoCs. Drop-in models for an MCU project where running a full TensorFlow runtime is out of the question.

Models included

Model	Task	Input	Params	Int8 size	Latency (ESP32-S3 @ 240MHz)
kws_micro_v2	Keyword spotting (8 words)	49 × 10 MFCC	23k	31 KB	12 ms
person_detect_v3	Person yes/no in 96×96 RGB	96 × 96 × 3	250k	308 KB	95 ms
gesture_5class_v1	Hand gesture (open/fist/point/swipe/ok)	64 × 64 × 1	41k	52 KB	38 ms
anomaly_ae_v1	1-D sensor reconstruction error → anomaly score	128 × 1	8k	14 KB	6 ms
vad_micro_v2	Voice activity 0/1 on 30 ms windows	24 MFCC	6k	11 KB	4 ms
defect_binary_v1	Defect yes/no on 64×64 grayscale	64 × 64 × 1	38k	48 KB	32 ms

Each model ships in three formats:

• TFLite (int8) — for TensorFlow Lite Micro on Cortex-M, ESP32 via ESP-NN, Raspberry Pi
• ONNX (int8 + fp16) — for ONNX Runtime on Pi or any board with x86/ARM64
• ESP-DL — Espressif's optimized format for ESP32-S3's vector ISA. Roughly 2× faster than generic TFLite on S3.

Latency table (all platforms)

Model	Cortex-M4F @ 80MHz	ESP32-S3 @ 240MHz (TFLite)	ESP32-S3 @ 240MHz (ESP-DL)	RPi 4 (TFLite)
kws_micro_v2	38 ms	12 ms	7 ms	3 ms
person_detect_v3	n/a (RAM)	95 ms	58 ms	14 ms
gesture_5class_v1	121 ms	38 ms	19 ms	6 ms
anomaly_ae_v1	18 ms	6 ms	3 ms	1 ms
vad_micro_v2	12 ms	4 ms	2 ms	1 ms
defect_binary_v1	105 ms	32 ms	16 ms	5 ms

"n/a (RAM)" = model's working memory exceeds the SoC's tightly-coupled SRAM budget.

Why pre-quantized

Int8 post-training quantization typically costs 0.5–2% accuracy but is essential for fitting in MCU SRAM. We did per-channel quantization with a 1000-sample calibration set per model and verified accuracy drop is ≤2% vs the float baseline:

Model	FP32 accuracy	INT8 accuracy	Δ
kws_micro_v2	94.2%	93.4%	-0.8%
person_detect_v3	89.1%	87.6%	-1.5%
gesture_5class_v1	96.8%	95.9%	-0.9%
anomaly_ae_v1	F1=0.91	F1=0.89	-0.02
vad_micro_v2	97.5%	97.1%	-0.4%
defect_binary_v1	91.4%	89.9%	-1.5%

Integration examples

examples/esp32_kws.cpp and examples/rpi_person_detect.py show end-to-end inference loops on the two most common targets. Audio + camera input adapters are sketched but use the platform's native APIs — those parts you swap for your own peripheral driver.

License

MIT. The models themselves are derived works of public-domain datasets (Google Speech Commands, COCO person subset, our own synthetic defect dataset) and may be redistributed under MIT.