Synaptics/yolov26n_od

YOLO26 320x320 (nano and small)

Model Overview

YOLO26 object detection models optimized for 320x320 resolution — Ultralytics' end-to-end NMS-free architecture with the decode head compiled onto the NPU. This repo carries both the nano (yolo26n) and small (yolo26s) variants.

Both models are developed and optimized for the Synaptics Astra™ SL2610-Series processors Torq NPU.

Model Features

• Model Type: Object Detection (NMS-free, one-to-one head)
• Input: 1x320x320x3 int8 NHWC (scale 1/255, zero-point −128)
• Output: 1x84x2100 int8 — rows 0-3: xywh normalized to imgsz; rows 4-83: class probabilities (already decoded on-NPU; int8 output quantization saturates confidences near 0.5, threshold accordingly)
  • yolo26n output dequantization: scale 0.00423651235178113, zero-point −128
  • yolo26s output dequantization: scale 0.004613031819462776, zero-point −128

Measured performance (SL2610 NPU, iree-benchmark-module, 10 reps)

Model	Latency	fps	COCO val2017 mAP50 (int8)
yolo26n (yolo26n_npu.vmfb)	12.9 ms	78	0.403
yolo26s (yolo26s_npu.vmfb)	33.5 ms	30	0.503

Post-processing

Dequantize the output, threshold max class probability (~0.25), scale xywh by the input size, undo the letterbox. The NMS-free head needs no score-based NMS, but int8 confidence saturation produces tied duplicate boxes — apply an IoU dedupe pass.

Files

• yolo26n_full_integer_quant_320_od.tflite / yolo26s_full_integer_quant_320_od.tflite — quantized source models (Ultralytics static-INT8 export, boundary quantize/dequantize stripped for int8 I/O)
• yolo26n_npu.vmfb / yolo26s_npu.vmfb — compiled Torq NPU binaries
• yolo26n_metadata_od.yaml / yolo26s_metadata_od.yaml — export metadata (class names, I/O description)
• labels.json, samples/ — demo assets

The hosted vmfbs are built with torq-compile from main (the YOLO26 compiler fixes, synaptics-torq/torq-compiler-dev #2285 and #2337, are merged).

License

Both the source models and the compiled models for on-device deployment are licensed under AGPL-3.0.