LibreYOLO Object Detection Leaderboard

Every detection model in LibreYOLO, measured on one protocol across hardware and runtimes. Pick the best one for your use case.

Benchmark runs708

RF100-VL: fine-tuned transfer

RF100-VL asks a different question than COCO: not how well a pre-trained checkpoint scores, but how well an architecture adapts. Each of the 100 Roboflow Universe datasets is fine-tuned separately and scored on its own test split with pycocotools at maxDets 500; the headline number is the unweighted mean across those 100 scores.

These values are not comparable to the COCO mAP in the leaderboard above, and not comparable to each other across a different protocol. Latency is not reported here because a run spans 100 separate checkpoints.

Measured on NVIDIA RTX 5060 Ti · PyTorch FP32.

Model	AP@50-95	AP@50	Params (M)	GFLOPs	Datasets
yolov9s	55.9%	81.4%	7.2M	13.5	100/100
yolov9t	54.0%	79.6%	2.0M	4.0	100/100

RF100-VL AP@50-95 vs parameter count

How much transfer accuracy each model buys per parameter.

yolov9

Benchmark data tables

Accuracy vs model size for 78 models across 16 families. Highest accuracy: deimv2-x at 61.3 mAP@50-95 (51.2 M params). Lightest is deimv2-atto at 0.5 M params (27.5 mAP).
Model	Family	mAP@50-95 (%)	Params (M)	GFLOPs
deimv2-x	deimv2	61.3	51.2	151.6
ec-x	ec	61.1	49.9	151.0
ec-l	ec	60.1	33.0	101.0
rtdetrv4-x	rtdetrv4	60.0	62.6	202.0
deim-x	deim	59.6	62.6	202.0
dfine-x	dfine	59.3	62.6	202.0
deimv2-l	deimv2	58.6	32.5	96.3
rfdetr-l	rfdetr	58.5	33.9	340.0
ec-m	ec	58.4	19.4	53.0
rtdetr-x	rtdetr	57.9	67.4	234.0
deim-l	deim	57.8	31.2	91.0
rtdetrv4-l	rtdetrv4	57.8	31.2	91.0
rfdetr-m	rfdetr	57.4	33.7	0.0
dfine-l	dfine	57.3	31.2	91.0
yolov9c	yolov9	57.1	25.5	51.8
rtdetrv2-r101	rtdetrv2	56.8	76.6	259.0
rtdetrv4-m	rtdetrv4	56.5	19.6	57.0
yolo26x	yolo26	56.5	55.7	193.9
yolox-x	yolox	56.3	99.1	141.2
yolov9m	yolov9	56.1	20.1	38.7
deimv2-m	deimv2	56.0	18.4	52.2
rtdetr-l	rtdetr	55.8	32.9	110.0
rtdetrv2-r50	rtdetrv2	55.7	42.9	136.0
deim-m	deim	55.4	19.6	57.0
yolox-l	yolox	55.4	54.2	78.0
rfdetr-s	rfdetr	55.1	32.1	0.0
dfine-m	dfine	55.1	19.6	57.0
rtdetrv2-r50m	rtdetrv2	54.8	36.6	100.0
ec-s	ec	54.3	9.9	26.0
rtdetr-r101	rtdetr	53.9	76.6	259.0
yolo26l	yolo26	53.8	24.8	86.4
yolo11x	yolo11	53.6	56.9	194.9
rtdetrv2-r34	rtdetrv2	53.2	31.4	92.0
yolov8x	yolov8	53.0	68.2	257.8
deimv2-s	deimv2	53.0	9.8	25.6
rtdetrv4-s	rtdetrv4	52.8	10.3	25.0
rtdetr-r50	rtdetr	52.7	42.9	136.0
yolov5xu	yolov5u	52.3	97.2	246.4
yolo11l	yolo11	52.3	25.3	86.9
yolo26m	yolo26	52.2	20.4	68.2
deim-s	deim	52.1	10.3	25.0
yolov8l	yolov8	52.0	43.7	165.2
yolox-m	yolox	51.7	25.3	37.0
yolov5lu	yolov5u	51.5	53.2	135.0
rfdetr-n	rfdetr	51.4	30.5	0.0
yolonas-l	yolonas	51.2	67.0	116.6
rtdetr-r50m	rtdetr	50.8	36.6	0.0
rtdetrv2-r18	rtdetrv2	50.8	20.2	60.0
dfine-s	dfine	50.7	10.3	25.0
yolo11m	yolo11	50.6	20.1	68.0
yolonas-m	yolonas	50.5	51.2	88.9
yolov9s	yolov9	50.5	7.2	13.5
yolov8m	yolov8	49.4	25.9	78.9
rtdetr-r34	rtdetr	48.2	31.4	91.0
yolov5mu	yolov5u	48.2	25.1	64.2
yolo26s	yolo26	47.4	9.5	20.7
deim-n	deim	46.8	3.8	7.0
deimv2-n	deimv2	46.7	3.6	6.9
yolonas-s	yolonas	46.5	19.1	32.8
yolo11s	yolo11	45.7	9.4	21.5
rtdetr-r18	rtdetr	45.6	20.2	60.0
yolox-s	yolox	44.3	9.0	13.5
yolov8s	yolov8	44.2	11.2	28.6
picodet-l	picodet	44.1	3.3	8.9
dfine-n	dfine	42.8	3.8	7.0
yolov5su	yolov5u	42.4	9.1	24.0
deimv2-pico	deimv2	42.2	1.5	5.2
yolov9t	yolov9	41.8	2.0	4.0
yolo26n	yolo26	40.1	2.4	5.4
yolo11n	yolo11	38.7	2.6	6.5
picodet-m	picodet	37.9	2.1	2.5
yolov8n	yolov8	36.7	3.1	8.7
yolox-tiny	yolox	35.5	5.1	7.7
deimv2-femto	deimv2	34.5	1.0	1.7
yolov5nu	yolov5u	33.8	2.6	7.7
picodet-s	picodet	30.4	1.0	0.7
yolox-nano	yolox	28.8	0.9	1.3
deimv2-atto	deimv2	27.5	0.5	0.8

Accuracy vs latency on NVIDIA A100 · PyTorch FP32 for 13 models across 3 families. Highest accuracy: dfine-x at 59.3 mAP@50-95. Fastest is rtdetr-r34 at 36.0 ms (27.8 FPS).
Model	Family	mAP@50-95 (%)	Latency (ms)	FPS	Params (M)
dfine-x	dfine	59.3	64.7	15.4	62.6
dfine-l	dfine	57.3	58.3	17.1	31.2
dfine-m	dfine	55.1	46.1	21.7	19.6
rtdetr-r101	rtdetr	53.9	53.6	18.6	76.6
rtdetr-r50	rtdetr	52.7	44.1	22.7	42.9
yolonas-l	yolonas	51.2	60.4	16.6	67.0
rtdetr-r50m	rtdetr	50.8	39.8	25.1	36.6
dfine-s	dfine	50.7	41.7	24.0	10.3
yolonas-m	yolonas	50.5	58.9	17.0	51.2
rtdetr-r34	rtdetr	48.2	36.0	27.8	31.4
yolonas-s	yolonas	46.5	60.1	16.6	19.1
rtdetr-r18	rtdetr	45.6	36.5	27.4	20.2
dfine-n	dfine	42.8	39.1	25.6	3.8

Accuracy vs latency on NVIDIA Jetson Orin Nano Super 8GB · PyTorch FP32 for 78 models across 16 families. Highest accuracy: dfine-x at 61.4 mAP@50-95. Fastest is yolov8n at 37.8 ms (26.4 FPS).
Model	Family	mAP@50-95 (%)	Latency (ms)	FPS	Params (M)
dfine-x	dfine	61.4	350.0	2.9	62.6
deimv2-x	deimv2	61.3	370.1	2.7	51.2
ec-x	ec	61.1	349.5	2.9	49.9
ec-l	ec	60.1	296.0	3.4	33.0
rtdetrv4-x	rtdetrv4	60.0	351.8	2.8	62.6
dfine-l	dfine	60.0	211.7	4.7	31.2
deim-x	deim	59.6	349.8	2.9	62.6
deimv2-l	deimv2	58.6	303.8	3.3	32.5
rfdetr-l	rfdetr	58.5	248.5	4.0	33.9
ec-m	ec	58.4	200.8	5.0	19.4
rtdetr-x	rtdetr	57.9	338.1	3.0	67.4
deim-l	deim	57.8	209.3	4.8	31.2
dfine-m	dfine	57.8	143.3	7.0	19.6
rtdetrv4-l	rtdetrv4	57.8	210.8	4.7	31.2
rfdetr-m	rfdetr	57.4	174.5	5.7	33.7
yolov9c	yolov9	57.1	155.6	6.4	25.5
rtdetr-r101	rtdetr	56.8	359.6	2.8	76.6
rtdetrv2-r101	rtdetrv2	56.8	358.7	2.8	76.6
rtdetrv4-m	rtdetrv4	56.5	143.1	7.0	19.6
yolo26x	yolo26	56.5	193.6	5.2	55.7
yolonas-l	yolonas	56.3	200.2	5.0	67.0
yolox-x	yolox	56.3	289.6	3.5	99.1
yolov9m	yolov9	56.1	124.8	8.0	20.1
deimv2-m	deimv2	56.0	213.5	4.7	18.4
rtdetr-r50	rtdetr	55.9	234.6	4.3	42.9
rtdetr-l	rtdetr	55.8	206.0	4.9	32.9
rtdetrv2-r50	rtdetrv2	55.7	233.4	4.3	42.9
yolonas-m	yolonas	55.5	152.0	6.6	51.2
deim-m	deim	55.4	143.7	7.0	19.6
yolox-l	yolox	55.4	171.7	5.8	54.2
rfdetr-s	rfdetr	55.1	144.6	6.9	32.1
rtdetrv2-r50m	rtdetrv2	54.8	196.8	5.1	36.6
ec-s	ec	54.3	168.2	6.0	9.9
rtdetr-r50m	rtdetr	53.8	198.3	5.0	36.6
yolo26l	yolo26	53.8	107.2	9.3	24.8
yolo11x	yolo11	53.6	196.1	5.1	56.9
dfine-s	dfine	53.4	96.8	10.3	10.3
rtdetrv2-r34	rtdetrv2	53.2	123.1	8.1	31.4
yolov8x	yolov8	53.0	184.8	5.4	68.2
deimv2-s	deimv2	53.0	175.8	5.7	9.8
rtdetrv4-s	rtdetrv4	52.8	99.3	10.1	10.3
yolov5xu	yolov5u	52.3	200.2	5.0	97.2
yolo11l	yolo11	52.3	109.9	9.1	25.3
yolo26m	yolo26	52.2	88.6	11.3	20.4
rtdetr-r34	rtdetr	52.2	125.8	8.0	31.4
deim-s	deim	52.1	98.1	10.2	10.3
yolov8l	yolov8	52.0	120.3	8.3	43.7
yolonas-s	yolonas	51.8	98.0	10.2	19.1
yolox-m	yolox	51.7	110.1	9.1	25.3
yolov5lu	yolov5u	51.5	116.9	8.6	53.2
rfdetr-n	rfdetr	51.4	98.3	10.2	30.5
rtdetrv2-r18	rtdetrv2	50.8	97.4	10.3	20.2
yolo11m	yolo11	50.6	90.3	11.1	20.1
yolov9s	yolov9	50.5	101.2	9.9	7.2
rtdetr-r18	rtdetr	49.8	98.3	10.2	20.2
yolov8m	yolov8	49.4	82.1	12.2	25.9
yolov5mu	yolov5u	48.2	77.9	12.8	25.1
yolo26s	yolo26	47.4	49.9	20.1	9.5
deim-n	deim	46.8	88.5	11.3	3.8
deimv2-n	deimv2	46.7	89.8	11.1	3.6
dfine-n	dfine	45.8	85.5	11.7	3.8
yolo11s	yolo11	45.7	48.3	20.7	9.4
yolox-s	yolox	44.3	59.2	16.9	9.0
yolov8s	yolov8	44.2	45.1	22.1	11.2
picodet-l	picodet	44.1	90.1	11.1	3.3
yolov5su	yolov5u	42.4	44.0	22.8	9.1
deimv2-pico	deimv2	42.2	71.1	14.1	1.5
yolov9t	yolov9	41.8	101.2	9.9	2.0
yolo26n	yolo26	40.1	46.4	21.6	2.4
yolo11n	yolo11	38.7	43.8	22.9	2.6
picodet-m	picodet	37.9	80.3	12.5	2.1
yolov8n	yolov8	36.7	37.8	26.4	3.1
yolox-tiny	yolox	35.5	49.2	20.3	5.1
deimv2-femto	deimv2	34.5	67.5	14.8	1.0
yolov5nu	yolov5u	33.8	38.3	26.1	2.6
picodet-s	picodet	30.4	72.2	13.8	1.0
yolox-nano	yolox	28.8	54.5	18.4	0.9
deimv2-atto	deimv2	27.5	64.9	15.4	0.5

Accuracy vs latency on NVIDIA Jetson Orin Nano Super 8GB · ONNX Runtime FP32 for 75 models across 15 families. Highest accuracy: dfine-x at 61.4 mAP@50-95. Fastest is deimv2-atto at 22.8 ms (43.9 FPS).
Model	Family	mAP@50-95 (%)	Latency (ms)	FPS	Params (M)
dfine-x	dfine	61.4	328.9	3.0	61.9
deimv2-x	deimv2	61.3	444.9	2.3	50.3
ec-x	ec	61.1	422.4	2.4	49.3
ec-l	ec	60.1	367.3	2.7	33.0
rtdetrv4-x	rtdetrv4	60.0	342.2	2.9	62.8
dfine-l	dfine	60.0	199.6	5.0	30.9
deim-x	deim	59.6	328.8	3.0	61.9
rfdetr-l	rfdetr	58.6	295.5	3.4	30.4
deimv2-l	deimv2	58.5	370.2	2.7	32.2
ec-m	ec	58.4	254.4	3.9	19.5
rtdetr-x	rtdetr	58.0	331.4	3.0	67.5
deim-l	deim	57.8	199.7	5.0	30.9
dfine-m	dfine	57.8	144.4	6.9	19.4
rtdetrv4-l	rtdetrv4	57.8	206.5	4.8	31.3
rfdetr-m	rfdetr	57.4	204.7	4.9	30.1
yolov9c	yolov9	57.1	156.1	6.4	25.5
rtdetr-r101	rtdetr	56.8	343.9	2.9	76.7
rtdetrv2-r101	rtdetrv2	56.8	343.3	2.9	76.7
rtdetrv4-m	rtdetrv4	56.5	149.1	6.7	19.6
yolo26x	yolo26	56.5	1767.7	0.6	55.7
yolox-x	yolox	56.3	273.8	3.6	99.0
yolov9m	yolov9	56.1	128.1	7.8	20.1
deimv2-m	deimv2	56.0	260.4	3.8	18.4
rtdetr-r50	rtdetr	55.9	228.0	4.4	43.0
rtdetr-l	rtdetr	55.8	206.5	4.8	33.0
rtdetrv2-r50	rtdetrv2	55.7	227.4	4.4	43.0
yolox-l	yolox	55.5	161.9	6.2	54.2
deim-m	deim	55.4	144.6	6.9	19.4
rfdetr-s	rfdetr	55.1	161.5	6.2	28.5
rtdetrv2-r50m	rtdetrv2	54.8	184.3	5.4	33.3
ec-s	ec	54.3	192.3	5.2	10.1
rtdetr-r50m	rtdetr	53.8	184.5	5.4	33.3
yolo26l	yolo26	53.8	855.2	1.2	24.8
yolo11x	yolo11	53.6	1787.0	0.6	56.9
dfine-s	dfine	53.4	96.1	10.4	10.3
rtdetrv2-r34	rtdetrv2	53.2	131.5	7.6	31.5
yolov8x	yolov8	53.0	2188.6	0.5	68.2
deimv2-s	deimv2	53.0	194.4	5.1	10.0
rtdetrv4-s	rtdetrv4	52.8	97.5	10.3	10.4
yolov5xu	yolov5u	52.3	2179.7	0.5	97.2
yolo11l	yolo11	52.3	874.4	1.1	25.3
rtdetr-r34	rtdetr	52.2	131.7	7.6	31.5
yolo26m	yolo26	52.2	674.0	1.5	20.4
deim-s	deim	52.1	96.2	10.4	10.3
yolov8l	yolov8	52.0	1451.2	0.7	43.7
yolox-m	yolox	51.8	107.3	9.3	25.3
yolov5lu	yolov5u	51.5	1247.2	0.8	53.2
rfdetr-n	rfdetr	51.4	96.0	10.4	26.9
rtdetrv2-r18	rtdetrv2	50.8	103.5	9.7	20.3
yolo11m	yolo11	50.6	686.2	1.5	20.1
yolov9s	yolov9	50.4	79.8	12.5	7.2
rtdetr-r18	rtdetr	49.8	103.7	9.6	20.3
yolov8m	yolov8	49.4	740.5	1.4	25.9
yolov5mu	yolov5u	48.2	632.2	1.6	25.1
yolo26s	yolo26	47.4	247.5	4.0	9.5
deim-n	deim	46.8	57.2	17.5	3.8
deimv2-n	deimv2	46.7	56.8	17.6	3.6
dfine-n	dfine	45.8	57.0	17.5	3.8
yolo11s	yolo11	45.7	265.9	3.8	9.4
yolox-s	yolox	44.3	58.6	17.1	9.0
yolov8s	yolov8	44.2	313.3	3.2	11.2
picodet-l	picodet	43.6	232.1	4.3	3.3
yolov5su	yolov5u	42.4	275.9	3.6	9.1
deimv2-pico	deimv2	42.3	47.1	21.3	1.5
yolov9t	yolov9	41.8	53.8	18.6	2.0
yolo26n	yolo26	40.1	96.5	10.4	2.4
yolo11n	yolo11	38.7	121.5	8.2	2.6
picodet-m	picodet	37.3	93.4	10.7	2.1
yolov8n	yolov8	36.7	137.1	7.3	3.2
yolox-tiny	yolox	35.5	31.8	31.4	5.0
deimv2-femto	deimv2	34.5	26.9	37.1	1.0
yolov5nu	yolov5u	33.8	127.1	7.9	2.6
picodet-s	picodet	29.6	61.1	16.4	1.0
yolox-nano	yolox	28.8	29.1	34.4	0.9
deimv2-atto	deimv2	27.5	22.8	43.9	0.5

Accuracy vs latency on NVIDIA Jetson Orin Nano Super 8GB · TensorRT FP32 for 55 models across 11 families. Highest accuracy: dfine-x at 61.5 mAP@50-95. Fastest is deimv2-atto at 12.3 ms (81.4 FPS).
Model	Family	mAP@50-95 (%)	Latency (ms)	FPS	Params (M)
dfine-x	dfine	61.5	124.8	8.0	62.0
deimv2-x	deimv2	61.3	194.2	5.2	50.3
ec-x	ec	61.1	186.5	5.4	49.0
ec-l	ec	60.1	163.8	6.1	31.0
rtdetrv4-x	rtdetrv4	60.0	128.3	7.8	62.0
dfine-l	dfine	60.0	80.2	12.5	31.0
deim-x	deim	59.6	125.1	8.0	62.0
rfdetr-l	rfdetr	58.6	127.8	7.8	128.0
deimv2-l	deimv2	58.6	163.3	6.1	32.2
ec-m	ec	58.3	121.6	8.2	18.0
rtdetr-x	rtdetr	58.0	115.8	8.6	67.0
deim-l	deim	57.8	80.5	12.4	31.0
rtdetrv4-l	rtdetrv4	57.8	82.4	12.1	31.0
dfine-m	dfine	57.7	63.5	15.8	19.0
rfdetr-m	rfdetr	57.4	85.9	11.6	33.7
yolov9c	yolov9	57.1	76.0	13.2	25.5
rtdetr-r101	rtdetr	56.8	118.5	8.4	76.0
rtdetrv2-r101	rtdetrv2	56.8	118.2	8.5	76.0
rtdetrv4-m	rtdetrv4	56.4	65.0	15.4	19.0
yolox-x	yolox	56.3	149.3	6.7	99.1
yolov9m	yolov9	56.1	68.9	14.5	20.1
deimv2-m	deimv2	55.9	125.4	8.0	18.1
rtdetr-r50	rtdetr	55.9	79.1	12.6	42.0
rtdetr-l	rtdetr	55.8	73.5	13.6	32.0
rtdetrv2-r50	rtdetrv2	55.7	78.6	12.7	42.0
deim-m	deim	55.5	63.8	15.7	19.0
yolox-l	yolox	55.4	87.5	11.4	54.2
rfdetr-s	rfdetr	55.1	69.1	14.5	32.1
rtdetrv2-r50m	rtdetrv2	54.8	63.3	15.8	36.0
ec-s	ec	54.4	94.8	10.6	9.9
rtdetr-r50m	rtdetr	53.8	63.6	15.7	36.6
dfine-s	dfine	53.4	44.7	22.4	10.0
rtdetrv2-r34	rtdetrv2	53.2	58.0	17.3	31.0
deimv2-s	deimv2	53.0	95.7	10.4	9.7
rtdetrv4-s	rtdetrv4	52.9	45.3	22.1	10.0
rtdetr-r34	rtdetr	52.2	58.5	17.1	31.0
deim-s	deim	52.1	45.0	22.2	10.0
yolox-m	yolox	51.8	61.1	16.4	25.3
rfdetr-n	rfdetr	51.4	42.1	23.7	30.5
rtdetrv2-r18	rtdetrv2	50.7	44.8	22.3	20.0
yolov9s	yolov9	50.5	46.2	21.6	7.2
rtdetr-r18	rtdetr	49.8	45.3	22.1	20.0
deim-n	deim	46.8	30.0	33.4	4.0
deimv2-n	deimv2	46.7	30.3	33.0	3.6
dfine-n	dfine	45.8	29.9	33.5	4.0
yolox-s	yolox	44.3	38.3	26.1	9.0
picodet-l	picodet	44.1	60.8	16.4	3.3
deimv2-pico	deimv2	42.3	26.5	37.7	1.5
yolov9t	yolov9	41.8	36.0	27.8	2.0
picodet-m	picodet	37.9	33.7	29.6	2.1
yolox-tiny	yolox	35.5	23.0	43.6	5.1
deimv2-femto	deimv2	34.5	16.7	60.0	1.0
picodet-s	picodet	30.4	26.4	37.9	1.0
yolox-nano	yolox	28.8	21.1	47.4	0.9
deimv2-atto	deimv2	27.5	12.3	81.4	0.5

Accuracy vs latency on NVIDIA RTX 5070 Ti · ONNX Runtime FP32 for 55 models across 11 families. Highest accuracy: dfine-x at 61.4 mAP@50-95. Fastest is deimv2-atto at 9.8 ms (102.0 FPS).
Model	Family	mAP@50-95 (%)	Latency (ms)	FPS	Params (M)
dfine-x	dfine	61.4	24.1	41.5	61.7
deimv2-x	deimv2	61.4	36.9	27.1	50.3
ec-x	ec	61.1	31.3	31.9	49.0
ec-l	ec	60.1	29.8	33.6	32.7
rtdetrv4-x	rtdetrv4	60.0	26.1	38.3	62.6
dfine-l	dfine	60.0	19.6	50.9	30.8
deim-x	deim	59.6	25.1	39.8	61.7
rfdetr-l	rfdetr	58.6	22.8	43.8	30.3
deimv2-l	deimv2	58.6	31.9	31.4	32.2
ec-m	ec	58.4	24.9	40.2	19.2
rtdetr-x	rtdetr	58.0	20.7	48.4	67.3
deim-l	deim	57.8	19.8	50.5	30.8
rtdetrv4-l	rtdetrv4	57.8	20.8	48.0	31.2
dfine-m	dfine	57.8	16.7	59.8	19.3
rfdetr-m	rfdetr	57.4	16.7	59.8	30.1
yolov9c	yolov9	57.1	15.7	63.5	25.5
rtdetr-r101	rtdetr	56.8	21.1	47.5	76.5
rtdetrv2-r101	rtdetrv2	56.8	20.3	49.3	76.6
rtdetrv4-m	rtdetrv4	56.5	17.7	56.6	19.5
yolox-x	yolox	56.3	24.0	41.6	99.0
yolov9m	yolov9	56.1	15.5	64.5	20.1
deimv2-m	deimv2	56.0	27.7	36.0	18.1
rtdetr-r50	rtdetr	55.9	15.9	62.8	42.8
rtdetr-l	rtdetr	55.8	15.5	64.5	32.8
rtdetrv2-r50	rtdetrv2	55.6	15.2	66.0	42.9
deim-m	deim	55.5	17.4	57.6	19.3
yolox-l	yolox	55.4	19.8	50.6	54.2
rfdetr-s	rfdetr	55.1	14.6	68.5	28.5
rtdetrv2-r50m	rtdetrv2	54.8	12.6	79.6	33.2
ec-s	ec	54.3	23.4	42.8	9.8
rtdetr-r50m	rtdetr	53.8	13.1	76.4	33.1
dfine-s	dfine	53.4	14.1	70.9	10.3
rtdetrv2-r34	rtdetrv2	53.2	12.3	81.1	31.4
deimv2-s	deimv2	53.0	25.2	39.6	9.7
rtdetrv4-s	rtdetrv4	52.8	15.0	66.7	10.3
rtdetr-r34	rtdetr	52.2	12.9	77.7	31.3
deim-s	deim	52.1	14.8	67.5	10.3
yolox-m	yolox	51.8	17.9	55.8	25.3
rfdetr-n	rfdetr	51.4	15.2	65.7	26.9
rtdetrv2-r18	rtdetrv2	50.8	10.3	97.1	20.1
yolov9s	yolov9	50.4	18.3	54.5	7.2
rtdetr-r18	rtdetr	49.8	10.8	92.5	20.1
deim-n	deim	46.8	13.2	75.9	3.8
deimv2-n	deimv2	46.7	13.1	76.3	3.6
dfine-n	dfine	45.8	12.7	78.9	3.8
yolox-s	yolox	44.3	16.5	60.6	9.0
picodet-l	picodet	44.1	29.2	34.3	3.3
deimv2-pico	deimv2	42.3	12.5	80.2	1.5
yolov9t	yolov9	41.8	19.2	52.1	2.0
picodet-m	picodet	37.9	21.7	46.0	2.1
yolox-tiny	yolox	35.5	13.1	76.2	5.0
deimv2-femto	deimv2	34.5	11.1	90.3	1.0
picodet-s	picodet	30.4	17.9	55.8	1.0
yolox-nano	yolox	28.8	15.8	63.2	0.9
deimv2-atto	deimv2	27.5	9.8	102.0	0.5

Accuracy vs latency on NVIDIA RTX 5070 Ti · TensorRT FP16 for 51 models across 11 families. Highest accuracy: dfine-x at 61.5 mAP@50-95. Fastest is rtdetrv2-r34 at 7.5 ms (132.9 FPS).
Model	Family	mAP@50-95 (%)	Latency (ms)	FPS	Params (M)
dfine-x	dfine	61.5	13.7	73.0	62.0
ec-x	ec	61.0	24.2	41.3	49.0
deimv2-x	deimv2	60.8	26.4	37.9	50.3
dfine-l	dfine	60.0	14.4	69.3	31.0
ec-l	ec	60.0	24.0	41.7	31.0
rtdetrv4-x	rtdetrv4	60.0	14.1	71.0	62.0
deim-x	deim	59.5	14.1	70.7	62.0
rfdetr-l	rfdetr	58.6	21.0	47.6	128.0
deimv2-l	deimv2	58.5	25.9	38.7	32.2
ec-m	ec	58.2	20.5	48.9	18.0
rtdetr-x	rtdetr	57.9	12.2	82.3	67.0
rtdetrv4-l	rtdetrv4	57.7	12.8	78.0	31.0
dfine-m	dfine	57.6	13.4	74.5	19.0
rfdetr-m	rfdetr	57.4	15.9	62.9	33.7
yolov9c	yolov9	57.2	13.4	74.7	25.5
rtdetrv2-r101	rtdetrv2	56.8	11.4	87.4	76.0
rtdetr-r101	rtdetr	56.7	11.4	88.0	76.0
rtdetrv4-m	rtdetrv4	56.4	11.5	86.7	19.0
yolox-x	yolox	56.3	13.9	72.1	99.1
yolov9m	yolov9	56.1	13.8	72.3	20.1
rtdetr-l	rtdetr	55.8	8.2	122.0	32.0
rtdetr-r50	rtdetr	55.8	10.3	97.4	42.0
rtdetrv2-r50	rtdetrv2	55.7	9.7	103.6	42.0
yolox-l	yolox	55.4	12.6	79.3	54.2
deim-m	deim	55.4	12.0	83.3	19.0
rfdetr-s	rfdetr	55.3	14.0	71.6	32.1
rtdetrv2-r50m	rtdetrv2	54.7	8.9	113.0	36.0
ec-s	ec	54.4	23.0	43.5	9.9
rtdetr-r50m	rtdetr	53.9	9.1	110.1	36.6
dfine-s	dfine	53.2	11.6	86.5	10.0
rtdetrv2-r34	rtdetrv2	53.2	7.5	132.9	31.0
deimv2-s	deimv2	53.0	19.8	50.4	9.7
rtdetrv4-s	rtdetrv4	52.9	10.8	92.3	10.0
rtdetr-r34	rtdetr	52.3	9.2	108.4	31.0
deim-s	deim	51.9	12.1	82.4	10.0
yolox-m	yolox	51.6	12.0	83.0	25.3
rfdetr-n	rfdetr	51.4	17.1	58.3	30.5
rtdetrv2-r18	rtdetrv2	50.8	8.5	117.8	20.0
yolov9s	yolov9	50.5	15.7	63.8	7.2
rtdetr-r18	rtdetr	49.8	8.8	113.4	20.0
deimv2-n	deimv2	46.6	12.8	78.0	3.6
deim-n	deim	46.6	12.3	81.3	4.0
dfine-n	dfine	45.8	12.5	79.7	4.0
picodet-l	picodet	44.0	26.3	38.0	3.3
yolox-s	yolox	43.4	14.1	70.9	9.0
yolov9t	yolov9	41.8	17.0	59.0	2.0
picodet-m	picodet	37.3	15.2	65.7	2.1
yolox-tiny	yolox	35.4	10.2	97.7	5.1
picodet-s	picodet	30.4	12.7	78.6	1.0
yolox-nano	yolox	28.7	10.0	100.0	0.9
deimv2-atto	deimv2	25.8	7.6	132.1	0.5

Accuracy vs latency on NVIDIA RTX 5070 Ti · TensorRT FP32 for 55 models across 11 families. Highest accuracy: dfine-x at 61.4 mAP@50-95. Fastest is deimv2-atto at 7.3 ms (137.9 FPS).
Model	Family	mAP@50-95 (%)	Latency (ms)	FPS	Params (M)
dfine-x	dfine	61.4	17.3	57.9	62.0
deimv2-x	deimv2	61.3	26.2	38.1	50.3
ec-x	ec	61.1	21.9	45.6	49.0
ec-l	ec	60.1	20.4	49.0	31.0
rtdetrv4-x	rtdetrv4	60.0	17.1	58.3	62.0
dfine-l	dfine	60.0	13.8	72.3	31.0
deim-x	deim	59.6	17.1	58.6	62.0
deimv2-l	deimv2	58.6	23.2	43.2	32.2
rfdetr-l	rfdetr	58.6	22.2	45.1	128.0
ec-m	ec	58.4	16.9	59.2	18.0
rtdetr-x	rtdetr	57.9	14.6	68.5	67.0
deim-l	deim	57.9	13.2	75.7	31.0
rtdetrv4-l	rtdetrv4	57.8	13.8	72.7	31.0
dfine-m	dfine	57.8	11.7	85.3	19.0
rfdetr-m	rfdetr	57.4	15.7	63.7	33.7
yolov9c	yolov9	57.1	13.0	76.7	25.5
rtdetr-r101	rtdetr	56.8	15.5	64.4	76.0
rtdetrv2-r101	rtdetrv2	56.8	15.4	64.8	76.0
rtdetrv4-m	rtdetrv4	56.5	11.9	84.4	19.0
yolox-x	yolox	56.3	20.7	48.4	99.1
yolov9m	yolov9	56.1	12.9	77.5	20.1
deimv2-m	deimv2	56.0	20.2	49.6	18.1
rtdetr-r50	rtdetr	55.9	11.4	87.4	42.0
rtdetr-l	rtdetr	55.7	10.6	94.2	32.0
rtdetrv2-r50	rtdetrv2	55.7	11.3	88.5	42.0
deim-m	deim	55.5	11.7	85.4	19.0
yolox-l	yolox	55.4	16.2	61.8	54.2
rfdetr-s	rfdetr	55.1	12.9	77.5	32.1
rtdetrv2-r50m	rtdetrv2	54.8	9.7	103.3	36.0
ec-s	ec	54.4	15.8	63.4	9.9
rtdetr-r50m	rtdetr	53.9	9.7	102.8	36.6
dfine-s	dfine	53.4	10.1	98.7	10.0
rtdetrv2-r34	rtdetrv2	53.2	9.8	102.5	31.0
deimv2-s	deimv2	53.0	19.0	52.5	9.7
rtdetrv4-s	rtdetrv4	52.9	10.3	97.1	10.0
rtdetr-r34	rtdetr	52.2	10.2	98.1	31.0
deim-s	deim	52.1	10.2	97.6	10.0
yolox-m	yolox	51.7	14.3	70.1	25.3
rfdetr-n	rfdetr	51.4	14.8	67.6	30.5
rtdetrv2-r18	rtdetrv2	50.7	8.0	125.5	20.0
yolov9s	yolov9	50.5	13.3	75.1	7.2
rtdetr-r18	rtdetr	49.8	8.3	120.6	20.0
deim-n	deim	46.8	14.8	67.5	4.0
deimv2-n	deimv2	46.7	14.3	70.1	3.6
dfine-n	dfine	45.8	10.1	98.7	4.0
yolox-s	yolox	44.3	13.4	74.5	9.0
picodet-l	picodet	44.1	31.3	32.0	3.3
deimv2-pico	deimv2	42.2	12.5	79.9	1.5
yolov9t	yolov9	41.8	13.9	72.0	2.0
picodet-m	picodet	36.0	18.7	53.5	2.1
deimv2-femto	deimv2	34.3	8.4	119.0	1.0
yolox-tiny	yolox	33.9	11.5	87.2	5.1
picodet-s	picodet	30.3	13.5	74.0	1.0
yolox-nano	yolox	28.6	11.5	86.7	0.9
deimv2-atto	deimv2	27.5	7.3	137.9	0.5

Accuracy vs latency on NVIDIA Jetson Orin Nano Super 8GB · TensorRT FP16 for 75 models across 15 families. Highest accuracy: dfine-x at 61.4 mAP@50-95. Fastest is deimv2-atto at 11.2 ms (89.1 FPS).
Model	Family	mAP@50-95 (%)	Latency (ms)	FPS	Params (M)
dfine-x	dfine	61.4	70.5	14.2	62.0
ec-x	ec	61.0	174.8	5.7	49.0
deimv2-x	deimv2	60.8	184.5	5.4	50.3
ec-l	ec	60.0	161.5	6.2	31.0
dfine-l	dfine	60.0	53.6	18.6	31.0
rtdetrv4-x	rtdetrv4	60.0	72.7	13.8	62.0
deim-x	deim	59.6	70.9	14.1	62.0
rfdetr-l	rfdetr	58.7	119.3	8.4	128.0
deimv2-l	deimv2	58.5	160.8	6.2	32.2
ec-m	ec	58.3	119.7	8.3	18.0
dfine-m	dfine	57.9	42.3	23.7	19.0
deim-l	deim	57.9	54.0	18.5	31.0
rtdetr-x	rtdetr	57.9	58.9	17.0	67.0
rtdetrv4-l	rtdetrv4	57.8	55.2	18.1	31.0
rfdetr-m	rfdetr	57.4	79.8	12.5	33.7
yolov9c	yolov9	57.1	40.1	24.9	25.5
rtdetr-r101	rtdetr	56.8	58.8	17.0	76.0
rtdetrv2-r101	rtdetrv2	56.8	57.7	17.3	76.0
rtdetrv4-m	rtdetrv4	56.5	43.3	23.1	19.0
yolo26x	yolo26	56.5	58.9	17.0	55.7
yolox-x	yolox	56.2	68.7	14.6	99.1
yolov9m	yolov9	56.1	38.8	25.8	20.1
deimv2-m	deimv2	56.0	126.5	7.9	18.1
rtdetr-l	rtdetr	55.8	41.5	24.1	32.0
rtdetr-r50	rtdetr	55.7	42.4	23.6	42.0
rtdetrv2-r50	rtdetrv2	55.7	41.3	24.2	42.0
deim-m	deim	55.5	42.5	23.5	19.0
yolox-l	yolox	55.5	46.3	21.6	54.2
rfdetr-s	rfdetr	55.2	64.4	15.5	32.1
rtdetrv2-r50m	rtdetrv2	54.8	33.6	29.7	36.0
ec-s	ec	54.3	94.8	10.6	9.9
rtdetr-r50m	rtdetr	53.8	34.3	29.1	36.6
yolo26l	yolo26	53.8	41.2	24.3	24.8
yolo11x	yolo11	53.6	63.3	15.8	56.9
dfine-s	dfine	53.5	33.1	30.2	10.0
rtdetrv2-r34	rtdetrv2	53.2	30.6	32.7	31.0
yolov8x	yolov8	53.0	69.2	14.4	68.2
deimv2-s	deimv2	53.0	95.5	10.5	9.7
rtdetrv4-s	rtdetrv4	52.9	33.7	29.6	10.0
yolov5xu	yolov5u	52.3	68.1	14.7	97.2
rtdetr-r34	rtdetr	52.3	31.4	31.9	31.0
yolo11l	yolo11	52.3	44.3	22.6	25.3
yolo26m	yolo26	52.2	36.0	27.8	20.4
deim-s	deim	52.0	33.4	29.9	10.0
yolov8l	yolov8	52.0	51.9	19.3	43.7
yolox-m	yolox	51.8	35.7	28.0	25.3
yolov5lu	yolov5u	51.5	48.5	20.6	53.2
rfdetr-n	rfdetr	51.3	39.4	25.4	30.5
rtdetrv2-r18	rtdetrv2	50.8	25.1	39.9	20.0
yolo11m	yolo11	50.6	38.8	25.8	20.1
yolov9s	yolov9	50.4	30.1	33.2	7.2
rtdetr-r18	rtdetr	49.7	25.7	38.8	20.0
yolov8m	yolov8	49.4	40.6	24.6	25.9
yolov5mu	yolov5u	48.2	36.5	27.4	25.1
yolo26s	yolo26	47.4	24.0	41.7	9.5
deim-n	deim	46.8	23.9	41.8	4.0
deimv2-n	deimv2	46.6	24.3	41.2	3.6
dfine-n	dfine	45.8	23.8	42.1	4.0
yolo11s	yolo11	45.7	24.2	41.3	9.4
yolox-s	yolox	44.2	28.6	34.9	9.0
yolov8s	yolov8	44.2	25.2	39.7	11.2
picodet-l	picodet	44.0	52.3	19.1	3.3
yolov5su	yolov5u	42.4	23.4	42.8	9.1
deimv2-pico	deimv2	42.3	22.5	44.5	1.5
yolov9t	yolov9	41.8	29.1	34.3	2.0
yolo26n	yolo26	40.1	16.6	60.2	2.4
yolo11n	yolo11	38.7	19.0	52.6	2.6
picodet-m	picodet	37.9	30.8	32.4	2.1
yolov8n	yolov8	36.7	20.2	49.5	3.2
yolox-tiny	yolox	35.5	19.4	51.6	5.1
deimv2-femto	deimv2	34.5	15.2	65.7	1.0
yolov5nu	yolov5u	33.8	17.8	56.3	2.6
picodet-s	picodet	30.4	25.3	39.6	1.0
yolox-nano	yolox	28.8	19.6	51.0	0.9
deimv2-atto	deimv2	27.5	11.2	89.1	0.5

Accuracy vs latency on NVIDIA RTX 5070 Ti · PyTorch FP32 for 55 models across 11 families. Highest accuracy: dfine-x at 61.4 mAP@50-95. Fastest is yolox-tiny at 16.2 ms (61.6 FPS).
Model	Family	mAP@50-95 (%)	Latency (ms)	FPS	Params (M)
dfine-x	dfine	61.4	55.1	18.1	62.6
deimv2-x	deimv2	61.3	57.6	17.4	51.2
ec-x	ec	61.1	43.5	23.0	49.9
ec-l	ec	60.1	41.3	24.2	33.0
rtdetrv4-x	rtdetrv4	60.0	60.1	16.6	62.6
dfine-l	dfine	60.0	45.8	21.9	31.2
deim-x	deim	59.6	66.7	15.0	62.6
deimv2-l	deimv2	58.6	50.7	19.7	32.5
rfdetr-l	rfdetr	58.5	39.9	25.1	33.9
ec-m	ec	58.4	37.7	26.5	19.4
rtdetr-x	rtdetr	57.9	46.0	21.8	67.4
deim-l	deim	57.8	53.8	18.6	31.2
dfine-m	dfine	57.8	34.8	28.8	19.6
rtdetrv4-l	rtdetrv4	57.8	40.4	24.7	31.2
rfdetr-m	rfdetr	57.3	33.0	30.3	33.7
yolov9c	yolov9	57.1	23.6	42.3	25.5
rtdetr-r101	rtdetr	56.8	42.4	23.6	76.6
rtdetrv2-r101	rtdetrv2	56.8	39.0	25.7	76.6
rtdetrv4-m	rtdetrv4	56.5	29.8	33.6	19.6
yolox-x	yolox	56.3	25.2	39.6	99.1
yolov9m	yolov9	56.1	26.9	37.1	20.1
deimv2-m	deimv2	56.0	45.4	22.1	18.4
rtdetr-r50	rtdetr	55.9	33.6	29.8	42.9
rtdetr-l	rtdetr	55.8	36.1	27.7	32.9
rtdetrv2-r50	rtdetrv2	55.7	31.1	32.1	42.9
deim-m	deim	55.5	35.3	28.4	19.6
yolox-l	yolox	55.4	23.3	42.8	54.2
rfdetr-s	rfdetr	55.1	27.9	35.8	32.1
rtdetrv2-r50m	rtdetrv2	54.8	25.1	39.9	36.6
ec-s	ec	54.4	36.4	27.5	9.9
rtdetr-r50m	rtdetr	53.8	28.4	35.3	36.6
dfine-s	dfine	53.4	29.3	34.2	10.3
rtdetrv2-r34	rtdetrv2	53.2	24.8	40.4	31.4
deimv2-s	deimv2	53.0	41.2	24.3	9.8
rtdetrv4-s	rtdetrv4	52.9	25.1	39.8	10.3
rtdetr-r34	rtdetr	52.2	24.9	40.2	31.4
deim-s	deim	52.1	30.0	33.4	10.3
yolox-m	yolox	51.7	20.7	48.2	25.3
rfdetr-n	rfdetr	51.4	25.3	39.5	30.5
rtdetrv2-r18	rtdetrv2	50.8	19.6	51.1	20.2
yolov9s	yolov9	50.5	32.5	30.8	7.2
rtdetr-r18	rtdetr	49.8	21.2	47.2	20.2
deim-n	deim	46.8	27.0	37.0	3.8
deimv2-n	deimv2	46.7	28.5	35.1	3.6
dfine-n	dfine	45.8	30.7	32.5	3.8
yolox-s	yolox	44.3	20.0	50.0	9.0
picodet-l	picodet	44.1	25.3	39.5	3.3
deimv2-pico	deimv2	42.3	24.7	40.5	1.5
yolov9t	yolov9	41.8	31.3	31.9	2.0
picodet-m	picodet	37.9	22.2	45.1	2.1
yolox-tiny	yolox	35.5	16.2	61.6	5.1
deimv2-femto	deimv2	34.5	22.1	45.3	1.0
picodet-s	picodet	30.4	20.0	50.0	1.0
yolox-nano	yolox	28.8	18.7	53.6	0.9
deimv2-atto	deimv2	27.5	24.7	40.4	0.5

Accuracy vs latency on Raspberry Pi 5 · PyTorch FP32 for 70 models across 16 families. Highest accuracy: deimv2-x at 61.3 mAP@50-95. Fastest is deimv2-atto at 79.1 ms (12.7 FPS).
Model	Family	mAP@50-95 (%)	Latency (ms)	FPS	Params (M)
deimv2-x	deimv2	61.3	3711.8	0.3	51.2
ec-x	ec	61.1	3635.1	0.3	49.9
ec-l	ec	60.1	2850.3	0.3	33.0
dfine-l	dfine	60.0	2305.2	0.4	31.2
rfdetr-l	rfdetr	58.6	2110.2	0.5	33.9
deimv2-l	deimv2	58.6	2802.2	0.4	32.5
ec-m	ec	58.4	1868.7	0.5	19.4
deim-l	deim	57.8	2295.5	0.4	31.2
dfine-m	dfine	57.8	1577.2	0.6	19.6
rtdetrv4-l	rtdetrv4	57.8	2303.8	0.4	31.2
rfdetr-m	rfdetr	57.4	1281.0	0.8	33.7
rtdetrv4-m	rtdetrv4	56.5	1588.4	0.6	19.6
yolo26x	yolo26	56.5	2507.6	0.4	55.7
yolov9c	yolov9	56.5	2153.8	0.5	25.5
deimv2-m	deimv2	56.0	1913.2	0.5	18.4
rtdetr-r50	rtdetr	55.9	3317.6	0.3	42.9
rtdetr-l	rtdetr	55.8	2345.2	0.4	32.9
rtdetrv2-r50	rtdetrv2	55.7	3317.3	0.3	42.9
yolonas-m	yolonas	55.5	2028.6	0.5	51.2
deim-m	deim	55.4	1582.8	0.6	19.6
yolov9m	yolov9	55.3	1638.0	0.6	20.1
rfdetr-s	rfdetr	55.1	959.4	1.0	32.1
rtdetrv2-r50m	rtdetrv2	54.8	2810.3	0.4	36.6
ec-s	ec	54.3	1201.0	0.8	9.9
yolox-l	yolox	53.9	2627.0	0.4	54.2
rtdetr-r50m	rtdetr	53.8	2802.0	0.4	36.6
yolo26l	yolo26	53.8	1377.3	0.7	24.8
yolo11x	yolo11	53.6	2522.8	0.4	56.9
dfine-s	dfine	53.4	874.3	1.1	10.3
rtdetrv2-r34	rtdetrv2	53.2	1664.9	0.6	31.4
yolov8x	yolov8	53.0	2422.0	0.4	68.2
deimv2-s	deimv2	53.0	1190.9	0.8	9.8
rtdetrv4-s	rtdetrv4	52.8	879.9	1.1	10.3
yolov5xu	yolov5u	52.3	2484.4	0.4	97.2
yolo11l	yolo11	52.3	1385.3	0.7	25.3
rtdetr-r34	rtdetr	52.2	1669.0	0.6	31.4
yolo26m	yolo26	52.2	1108.3	0.9	20.4
deim-s	deim	52.1	878.2	1.1	10.3
yolov8l	yolov8	52.0	1762.9	0.6	43.7
yolonas-s	yolonas	51.8	1049.5	0.9	19.1
yolov5lu	yolov5u	51.5	1587.0	0.6	53.2
rfdetr-n	rfdetr	51.4	479.6	2.1	30.5
yolox-m	yolox	50.9	1284.1	0.8	25.3
rtdetrv2-r18	rtdetrv2	50.8	1244.5	0.8	20.2
yolo11m	yolo11	50.6	1104.3	0.9	20.1
rtdetr-r18	rtdetr	49.7	1245.4	0.8	20.2
yolov9s	yolov9	49.5	716.7	1.4	7.2
yolov8m	yolov8	49.4	906.5	1.1	25.9
yolov5mu	yolov5u	48.2	824.2	1.2	25.1
yolo26s	yolo26	47.4	441.8	2.3	9.5
deim-n	deim	46.8	431.0	2.3	3.8
deimv2-n	deimv2	46.7	423.0	2.4	3.6
dfine-n	dfine	45.8	429.1	2.3	3.8
yolo11s	yolo11	45.7	435.8	2.3	9.4
yolov8s	yolov8	44.2	421.1	2.4	11.2
picodet-l	picodet	44.2	540.3	1.9	3.3
yolox-s	yolox	43.0	626.6	1.6	9.0
yolov5su	yolov5u	42.4	397.4	2.5	9.1
deimv2-pico	deimv2	42.2	349.8	2.9	1.5
yolov9t	yolov9	41.4	346.3	2.9	2.0
yolo26n	yolo26	40.1	198.1	5.0	2.4
yolo11n	yolo11	38.7	204.7	4.9	2.6
picodet-m	picodet	37.6	201.1	5.0	2.1
yolov8n	yolov8	36.7	186.8	5.3	3.1
yolox-tiny	yolox	35.1	183.5	5.5	5.1
deimv2-femto	deimv2	34.5	135.3	7.4	1.0
yolov5nu	yolov5u	33.8	184.6	5.4	2.6
picodet-s	picodet	29.5	95.0	10.5	1.0
yolox-nano	yolox	28.8	117.0	8.6	0.9
deimv2-atto	deimv2	27.5	79.1	12.7	0.5

Accuracy vs latency on Raspberry Pi 5 · ONNX Runtime FP32 for 75 models across 15 families. Highest accuracy: dfine-x at 61.4 mAP@50-95. Fastest is deimv2-atto at 33.0 ms (30.3 FPS).
Model	Family	mAP@50-95 (%)	Latency (ms)	FPS	Params (M)
dfine-x	dfine	61.4	2638.7	0.4	61.7
deimv2-x	deimv2	61.3	3453.0	0.3	50.3
ec-x	ec	61.1	3331.4	0.3	49.0
ec-l	ec	60.1	2681.8	0.4	32.7
dfine-l	dfine	60.0	1402.9	0.7	30.8
rtdetrv4-x	rtdetrv4	60.0	2685.7	0.4	62.6
deim-x	deim	59.6	2635.6	0.4	61.7
rfdetr-l	rfdetr	58.6	1924.5	0.5	30.3
deimv2-l	deimv2	58.6	2679.0	0.4	32.2
ec-m	ec	58.4	1708.2	0.6	19.2
rtdetr-x	rtdetr	58.0	2801.5	0.4	67.3
deim-l	deim	57.8	1400.2	0.7	30.8
dfine-m	dfine	57.8	992.4	1.0	19.3
rtdetrv4-l	rtdetrv4	57.8	1427.4	0.7	31.2
rfdetr-m	rfdetr	57.4	1142.2	0.9	30.1
rtdetr-r101	rtdetr	56.7	3019.4	0.3	76.5
rtdetrv2-r101	rtdetrv2	56.7	3042.7	0.3	76.6
rtdetrv4-m	rtdetrv4	56.5	1006.5	1.0	19.5
yolo26x	yolo26	56.5	2488.5	0.4	55.7
deimv2-m	deimv2	56.0	1699.7	0.6	18.1
rtdetr-r50	rtdetr	55.9	1772.9	0.6	42.8
rtdetr-l	rtdetr	55.8	1461.0	0.7	32.8
rtdetrv2-r50	rtdetrv2	55.7	1785.3	0.6	42.9
yolov9c	yolov9	55.7	1398.3	0.7	25.5
deim-m	deim	55.4	990.7	1.0	19.3
rfdetr-s	rfdetr	55.1	850.1	1.2	28.5
yolox-x	yolox	54.9	3200.5	0.3	99.0
rtdetrv2-r50m	rtdetrv2	54.8	1326.5	0.8	33.2
ec-s	ec	54.3	1154.9	0.9	9.8
yolox-l	yolox	54.3	1882.4	0.5	54.2
rtdetr-r50m	rtdetr	53.8	1329.0	0.8	33.1
yolo26l	yolo26	53.8	1241.6	0.8	24.8
yolo11x	yolo11	53.6	2526.0	0.4	56.9
yolov9m	yolov9	53.6	1039.9	1.0	20.1
dfine-s	dfine	53.4	559.3	1.8	10.3
rtdetrv2-r34	rtdetrv2	53.2	1122.2	0.9	31.4
yolov8x	yolov8	53.0	2873.9	0.3	68.2
deimv2-s	deimv2	53.0	1162.8	0.9	9.7
rtdetrv4-s	rtdetrv4	52.8	562.5	1.8	10.3
yolov5xu	yolov5u	52.3	2837.2	0.3	97.2
yolo11l	yolo11	52.3	1261.7	0.8	25.3
rtdetr-r34	rtdetr	52.2	1118.7	0.9	31.3
yolo26m	yolo26	52.2	980.6	1.0	20.4
deim-s	deim	52.1	559.2	1.8	10.3
yolov8l	yolov8	52.0	1913.7	0.5	43.7
yolov5lu	yolov5u	51.5	1625.4	0.6	53.2
rfdetr-n	rfdetr	51.4	409.2	2.4	26.9
yolox-m	yolox	50.8	922.0	1.1	25.3
rtdetrv2-r18	rtdetrv2	50.8	791.2	1.3	20.1
yolo11m	yolo11	50.6	994.5	1.0	20.1
rtdetr-r18	rtdetr	49.8	786.6	1.3	20.1
yolov8m	yolov8	49.4	972.3	1.0	25.9
yolov9s	yolov9	48.5	425.1	2.4	7.2
yolov5mu	yolov5u	48.2	818.5	1.2	25.1
yolo26s	yolo26	47.4	360.4	2.8	9.5
deim-n	deim	46.8	229.7	4.3	3.8
deimv2-n	deimv2	46.7	226.3	4.4	3.6
dfine-n	dfine	45.8	229.7	4.3	3.8
yolo11s	yolo11	45.7	378.1	2.6	9.4
yolov8s	yolov8	44.2	414.2	2.4	11.2
yolox-s	yolox	43.1	393.5	2.5	9.0
picodet-l	picodet	42.7	589.5	1.7	3.3
yolov5su	yolov5u	42.4	362.4	2.8	9.1
deimv2-pico	deimv2	42.3	161.5	6.2	1.5
yolov9t	yolov9	40.8	170.0	5.9	2.0
yolo26n	yolo26	40.1	140.6	7.1	2.4
yolo11n	yolo11	38.7	162.2	6.2	2.6
yolov8n	yolov8	36.7	170.4	5.9	3.2
picodet-m	picodet	36.5	160.2	6.2	2.1
deimv2-femto	deimv2	34.5	59.8	16.7	1.0
yolox-tiny	yolox	34.3	106.7	9.4	5.0
yolov5nu	yolov5u	33.8	156.6	6.4	2.6
picodet-s	picodet	28.1	66.8	15.0	1.0
yolox-nano	yolox	27.7	47.0	21.3	0.9
deimv2-atto	deimv2	27.5	33.0	30.3	0.5

Accuracy vs latency on Raspberry Pi 5 · ncnn FP32 for 33 models across 7 families. Highest accuracy: yolo26x at 56.5 mAP@50-95. Fastest is yolox-nano at 32.6 ms (30.7 FPS).
Model	Family	mAP@50-95 (%)	Latency (ms)	FPS	Params (M)
yolo26x	yolo26	56.5	1295.2	0.8	55.7
yolov9c	yolov9	55.8	531.8	1.9	25.5
yolox-x	yolox	53.9	1122.8	0.9	99.1
yolo26l	yolo26	53.8	641.4	1.6	24.8
yolo11x	yolo11	53.6	1296.7	0.8	56.9
yolov9m	yolov9	53.6	426.5	2.3	20.1
yolov8x	yolov8	53.0	982.0	1.0	68.2
yolox-l	yolox	52.9	662.4	1.5	54.2
yolov5xu	yolov5u	52.3	977.7	1.0	97.2
yolo11l	yolo11	52.3	650.1	1.5	25.3
yolo26m	yolo26	52.2	538.7	1.9	20.4
yolov8l	yolov8	52.0	685.9	1.5	43.7
yolov5lu	yolov5u	51.5	608.2	1.6	53.2
yolo11m	yolo11	50.6	532.7	1.9	20.1
yolox-m	yolox	49.7	340.3	2.9	25.3
yolov8m	yolov8	49.4	363.1	2.8	25.9
yolov9s	yolov9	48.4	203.8	4.9	7.2
yolov5mu	yolov5u	48.2	311.9	3.2	25.1
yolo26s	yolo26	47.4	176.9	5.7	9.5
yolo11s	yolo11	45.7	174.5	5.7	9.4
yolov8s	yolov8	44.2	175.5	5.7	11.2
picodet-l	picodet	42.8	441.8	2.3	3.3
yolov5su	yolov5u	42.4	162.5	6.2	9.1
yolox-s	yolox	41.9	174.3	5.7	9.0
yolov9t	yolov9	40.6	91.0	11.0	2.0
yolo26n	yolo26	40.1	77.6	12.9	2.4
yolo11n	yolo11	38.7	81.7	12.2	2.6
yolov8n	yolov8	36.7	83.7	11.9	3.2
picodet-m	picodet	36.5	117.8	8.5	2.1
yolox-tiny	yolox	34.0	54.3	18.4	5.1
yolov5nu	yolov5u	33.8	84.0	11.9	2.6
picodet-s	picodet	27.7	53.3	18.8	1.0
yolox-nano	yolox	26.8	32.6	30.7	0.9

Accuracy vs latency on Raspberry Pi 5 + Hailo-8 · HailoRT INT8 for 10 models across 3 families. Highest accuracy: yolov9c at 54.8 mAP@50-95. Fastest is yolox-tiny at 5.1 ms (195.6 FPS).
Model	Family	mAP@50-95 (%)	Latency (ms)	FPS	Params (M)
yolov9c	yolov9	54.8	47.1	21.2	25.5
yolov9m	yolov9	53.3	33.2	30.1	20.1
yolonas-l	yolonas	52.4	64.9	15.4	67.0
yolonas-m	yolonas	47.4	46.0	21.7	51.1
yolonas-s	yolonas	44.5	25.2	39.7	19.0
yolox-m	yolox	43.7	28.8	34.8	25.3
yolox-s	yolox	41.1	10.4	95.8	9.0
yolox-tiny	yolox	33.5	5.1	195.6	5.1
yolov9s	yolov9	32.0	27.5	36.3	7.2
yolov9t	yolov9	25.8	13.1	76.5	2.0

Accuracy vs latency on Raspberry Pi 5 · HailoRT INT8 for 5 models across 1 families. Highest accuracy: yolov8x at 46.0 mAP@50-95. Fastest is yolov8n at 12.4 ms (80.5 FPS).
Model	Family	mAP@50-95 (%)	Latency (ms)	FPS	Params (M)
yolov8x	yolov8	46.0	93.7	10.7	68.2
yolov8l	yolov8	44.3	58.1	17.2	43.7
yolov8m	yolov8	42.4	37.4	26.7	25.9
yolov8s	yolov8	36.9	14.7	68.0	11.2
yolov8n	yolov8	29.3	12.4	80.5	3.1

VA v1 Score

The composite ranking is coming back, but it will stay unpublished until the reviewed submission set is broad enough to make the ranking credible.

HardwareNVIDIA A100

RuntimePyTorch FP32

D-FINERF-DETRRT-DETRDEIMYOLOX

Preview only

25 of 25 modelsi

Vision Analysis

Ultralytics(YOLO11, YOLOv8)

Roboflow(RF-DETR)

Tsinghua(YOLOv10)

Baidu(RT-DETR)

Academia Sinica(YOLOv9)

Megvii(YOLOX)

Coming soon

Composite ranking in progress

VA v1 Score Over Time

The historical timeline is returning as part of the same composite score rollout. The chart stays visible as a preview, but the live series is not published yet.

Ultralytics

Megvii

Academia Sinica

Tsinghua

Roboflow

Baidu

Open Source