Model Support Reference¶

Complete reference for all model families supported by YOLOmatic — architectures, benchmarks, task support, and selection guidance.

YOLOmatic Version: 5.0.0 Last Updated: May 28, 2026

Table of Contents¶

Quick Comparison
Benchmark Compatibility
RF-DETR
YOLO26
YOLOv12
YOLO11
YOLOv10
YOLOv9
YOLOv8
YOLOX
SAM 3.1
Detectron2
Model Selection Recommendations
Supported Tasks
Architecture Evolution
Performance Comparison Summary
Roboflow Upload Notes
References

Quick Comparison¶

Family	Best Variant mAP	Smallest (Params)	Edge Optimized	Production Ready
RF-DETR	60.1 (2XLarge)	30.5M (Nano)	Yes	Yes
YOLO26	57.5 (x)	2.4M (n)	Yes	Yes
YOLOv12	55.2 (x)	2.6M (n)	No	No
YOLO11	54.7 (x)	2.6M (n)	Yes	Yes
YOLOv10	54.4 (X)	—	Yes	Yes
YOLOv9	55.6 (e)	2.0M (t)	Yes	Yes
YOLOv8	53.9 (x)	3.2M (n)	Yes	Yes
YOLOX	51.1 (X)	9.0M (S)	Yes	Yes

Benchmark Compatibility¶

The Benchmark & Vector Analysis feature (Evaluate & Monitor -> Benchmark Models or uv run yolomatic-benchmark) supports Ultralytics YOLO checkpoints and exported model artifacts that can be loaded by ultralytics.YOLO, including .pt, ONNX, TensorRT .engine, TorchScript, OpenVINO, TFLite, and other standard Ultralytics export targets. Task type (detection vs. segmentation) is auto-detected at runtime.

Artifact Type	Benchmark Support
YOLO detection checkpoint/export	Supported - Bounding-box IoU metrics
YOLO segmentation checkpoint/export	Supported - Pixel-level mask IoU metrics
ONNX / TensorRT / OpenVINO / other Ultralytics exports	Supported when the local runtime dependencies can execute the export
RF-DETR (`.pth`)	Not yet supported — use YOLO weights
SAM 3.1 (HuggingFace)	Not yet supported — use dedicated SAM inference
Detectron2 (`.pth`)	Not yet supported

Annotation format: COCO JSON (_annotations.coco.json). Roboflow export with COCO preset is directly compatible. YOLOmatic also includes a Labelbox NDJSON to COCO converter to assist with preparing custom validation sets.

RF-DETR¶

Status: Supported Focus: Real-time transformer detection and segmentation Trainer: Native rfdetr (not Ultralytics)

RF-DETR configs use the native RF-DETR trainer. Fresh training instantiates the selected model class without a local checkpoint so RF-DETR can automatically download and cache the official pretrained weights. Fine-tuning passes the selected .pth checkpoint as pretrain_weights; resume flows pass it as resume.

Detection Variants¶

Model	Class	mAP 50-95	Latency T4 TensorRT (ms)	Params (M)	Resolution	License
RF-DETR-Nano	`RFDETRNano`	48.4	2.3	30.5	384	Apache-2.0
RF-DETR-Small	`RFDETRSmall`	53.0	3.5	32.1	512	Apache-2.0
RF-DETR-Medium	`RFDETRMedium`	54.7	4.4	33.7	576	Apache-2.0
RF-DETR-Large	`RFDETRLarge`	56.5	6.8	33.9	704	Apache-2.0
RF-DETR-XLarge	`RFDETRXLarge`	58.6	11.5	126.4	700	PML-1.0
RF-DETR-2XLarge	`RFDETR2XLarge`	60.1	17.2	126.9	880	PML-1.0

Segmentation Variants¶

YOLOmatic also exposes RF-DETR segmentation classes from Nano through 2XLarge. These use task-specific default resolutions from the RF-DETR package and output .pth checkpoints discoverable by the prediction and Roboflow deployment flows.

Model	Class	Resolution	License
RF-DETR-Seg-Nano	`RFDETRSegNano`	312	Apache-2.0
RF-DETR-Seg-Small	`RFDETRSegSmall`	384	Apache-2.0
RF-DETR-Seg-Medium	`RFDETRSegMedium`	432	Apache-2.0
RF-DETR-Seg-Large	`RFDETRSegLarge`	504	Apache-2.0
RF-DETR-Seg-XLarge	`RFDETRSegXLarge`	624	Apache-2.0
RF-DETR-Seg-2XLarge	`RFDETRSeg2XLarge`	768	Apache-2.0

RF-DETR also ships a preview keypoint/pose model, RFDETRKeypointPreview (RF-DETR-Keypoint, default resolution 640, Apache-2.0). Pose training requires COCO keypoint JSON; YOLO pose datasets are auto-converted to a cached COCO manifest.

RF-DETR Plus (XLarge, 2XLarge) detection models require the rfdetr[plus] dependency extra and use PML-1.0 model licensing.

YOLO26¶

Status: Latest (2026-01-14) Focus: Edge deployment and end-to-end NMS-free inference

Key Features¶

End-to-End NMS-Free Inference — predictions generated directly without post-processing NMS
43% Faster CPU Inference — optimized for edge computing
DFL Removal — simplifies export and expands edge compatibility
MuSGD Optimizer — hybrid SGD + Muon optimizer inspired by LLM training breakthroughs
ProgLoss + STAL — improved loss functions for better accuracy, especially on small objects
Task Support — Detection, Segmentation, Classification, Pose, OBB

Detection Performance (COCO val2017)¶

Model	mAP 50-95	Speed T4 TensorRT (ms)	Params (M)	FLOPs (B)
YOLO26n	40.9	1.7 ± 0.0	2.4	5.4
YOLO26s	48.6	2.5 ± 0.0	9.5	20.7
YOLO26m	53.1	4.7 ± 0.1	20.4	68.2
YOLO26l	55.0	6.2 ± 0.2	24.8	86.4
YOLO26x	57.5	11.8 ± 0.2	55.7	193.9

CPU ONNX speed benchmarks for YOLO26 have not been officially published. T4 TensorRT figures shown above.

Segmentation Performance (COCO val2017)¶

Model	mAP box 50-95	Speed CPU ONNX (ms)	Params (M)	FLOPs (B)
YOLO26n-seg	33.9	53.3 ± 0.5	2.7	9.1
YOLO26s-seg	40.0	118.4 ± 0.9	10.4	34.2
YOLO26m-seg	44.1	328.2 ± 2.4	23.6	121.5
YOLO26l-seg	45.5	387.0 ± 3.7	28.0	139.8
YOLO26x-seg	47.0	787.0 ± 6.8	62.8	313.5

Mask mAP benchmarks for YOLO26-seg have not been officially published separately.

Pose Performance (COCO Keypoints)¶

Model	mAPpose 50-95	Params (M)	FLOPs (B)
YOLO26n-pose	-	2.7	6.0
YOLO26s-pose	-	9.7	21.5
YOLO26m-pose	-	20.6	69.0
YOLO26l-pose	-	25.0	87.0
YOLO26x-pose	-	56.0	194.5

Official mAP benchmarks for YOLO26-pose variants have not been published yet.

YOLOv12¶

Status: Community Model (2025 Early) Focus: High accuracy through attention mechanisms

Note: Ultralytics recommends YOLO11 or YOLO26 for production workloads due to training instability and higher memory consumption.

Key Features¶

Area Attention Mechanism — efficient large receptive field processing
R-ELAN — Residual Efficient Layer Aggregation Networks for better feature aggregation
FlashAttention Support — optional memory-efficient attention implementation
Task Support — Detection, Segmentation, Classification, Pose, OBB

Detection Performance (COCO val2017)¶

Model	mAP 50-95	Speed TensorRT (ms)	Params (M)	FLOPs (B)
YOLO12n	40.6	1.64	2.6	6.5
YOLO12s	48.0	2.61	9.3	21.4
YOLO12m	52.5	4.86	20.2	67.5
YOLO12l	53.7	6.77	26.4	88.9
YOLO12x	55.2	11.79	59.1	199.0

Improvements vs Previous Versions¶

vs YOLOv10n: +2.1% mAP, -9% speed
vs YOLO11m: +1.0% mAP, -3% speed
vs RT-DETRv2s: +0.1% mAP, +42% speed

YOLO11¶

Status: Stable (2024-09-10) Focus: Balanced accuracy and performance

Key Features¶

Enhanced Feature Extraction — improved backbone and neck architecture
Optimized Efficiency — refined architectural designs and training pipelines
Greater Accuracy with Fewer Parameters — 22% fewer params than YOLOv8m with higher mAP
Edge Deployment Ready — seamlessly deployable across various environments
Task Support — Detection, Segmentation, Classification, Pose, OBB

Detection Performance (COCO val2017)¶

Model	mAP 50-95	Speed CPU (ms)	Speed TensorRT (ms)	Params (M)	FLOPs (B)
YOLO11n	39.5	56.1 ± 0.8	1.5 ± 0.0	2.6	6.5
YOLO11s	47.0	90.0 ± 1.2	2.5 ± 0.0	9.4	21.5
YOLO11m	51.5	183.2 ± 2.0	4.7 ± 0.1	20.1	68.0
YOLO11l	53.4	238.6 ± 1.4	6.2 ± 0.1	25.3	86.9
YOLO11x	54.7	462.8 ± 6.7	11.3 ± 0.2	56.9	194.9

Segmentation Performance (COCO val2017)¶

Model	mAP box 50-95	mAP mask 50-95	Params (M)	FLOPs (B)
YOLO11n-seg	38.9	32.0	2.9	10.4
YOLO11s-seg	46.6	38.8	10.1	35.5
YOLO11m-seg	51.5	41.5	22.4	123.3
YOLO11l-seg	53.4	42.9	27.6	142.2
YOLO11x-seg	54.7	43.8	62.1	319.9

Pose Performance (COCO Keypoints)¶

Model	mAPpose 50-95	Speed CPU ONNX (ms)	Speed T4 TensorRT (ms)	Params (M)	FLOPs (B)
YOLO11n-pose	50.0	52.4 ± 0.5	1.7 ± 0.0	2.9	7.6
YOLO11s-pose	58.9	90.5 ± 0.6	2.6 ± 0.0	9.9	23.2
YOLO11m-pose	64.9	187.3 ± 0.8	4.9 ± 0.1	20.9	71.7
YOLO11l-pose	66.1	247.7 ± 1.1	6.4 ± 0.1	26.2	90.7
YOLO11x-pose	69.5	488.0 ± 13.9	12.1 ± 0.2	58.8	203.3

YOLOv10¶

Status: Stable Focus: End-to-end NMS-free inference pioneer

Model	mAP 50-95	FLOPs (G)	Latency (ms)
YOLOv10-N	38.5	6.7	1.84
YOLOv10-S	46.3	21.6	2.49
YOLOv10-M	51.1	59.1	4.74
YOLOv10-B	52.5	92.0	5.74
YOLOv10-L	53.2	120.3	7.28
YOLOv10-X	54.4	160.4	10.70

Tasks: Detection only.

YOLOv9¶

Status: Stable Focus: Extended feature optimization

Model	mAP 50-95	mAP 50	Params (M)	FLOPs (B)
YOLOv9t	38.3	53.1	2.0	7.7
YOLOv9s	46.8	63.4	7.2	26.7
YOLOv9m	51.4	68.1	20.1	76.8
YOLOv9c	53.0	70.2	25.5	102.8
YOLOv9e	55.6	72.8	58.1	192.5

Tasks: Detection + Segmentation (t/s/m/c/e + seg variants).

YOLOv8¶

Status: Stable Focus: Production-tested stability

Detection Performance¶

Model	mAP 50-95	Speed CPU ONNX (ms)	Params (M)	FLOPs (B)
YOLOv8n	37.3	80.4	3.2	8.7
YOLOv8s	44.9	128.4	11.2	28.6
YOLOv8m	50.2	234.7	25.9	78.9
YOLOv8l	52.9	375.2	43.7	165.2
YOLOv8x	53.9	479.1	68.2	257.8

Segmentation Performance¶

Model	mAP box 50-95	mAP mask 50-95	Params (M)	FLOPs (B)
YOLOv8n-seg	36.7	30.5	3.4	12.6
YOLOv8s-seg	44.6	36.8	11.8	42.6
YOLOv8m-seg	49.9	40.8	27.3	110.2
YOLOv8l-seg	52.3	42.6	46.0	220.5
YOLOv8x-seg	53.4	43.4	71.8	344.1

Pose Performance (COCO Keypoints)¶

Model	mAPpose 50-95	Speed CPU ONNX (ms)	Speed A100 TensorRT (ms)	Params (M)	FLOPs (B)
YOLOv8n-pose	50.4	131.8	1.18	3.3	9.2
YOLOv8s-pose	60.0	233.2	1.42	11.6	30.2
YOLOv8m-pose	65.0	456.3	2.00	26.4	81.0
YOLOv8l-pose	67.6	784.5	2.59	44.4	168.6
YOLOv8x-pose	70.2	1607.1	3.73	69.4	263.2

Tasks: Detection, Segmentation, Classification, Pose, OBB.

YOLOX¶

Status: Stable Focus: Decoupled head architecture

Model	mAP 50-95	Params (M)	FLOPs (G)	FPS
YOLOX-S	40.5%	9.0	26.8	102
YOLOX-M	46.9%	25.3	73.8	81
YOLOX-L	49.7%	54.2	155.6	69
YOLOX-X	51.1%	99.1	281.9	58

Tasks: Detection only.

SAM 3.1¶

Status: Supported Focus: Open-vocabulary segmentation (Segment Anything Model) Provider: Meta / HuggingFace (gated model)

Model	HuggingFace ID	Params (B)	Description
SAM 3.1	`facebook/sam3.1`	0.873	Object Multiplex — 7× faster multi-object throughput
SAM 3	`facebook/sam3`	0.848	Base model — predecessor without Object Multiplex

Capabilities: - Auto-segmentation (segment everything) - Text-prompted open-vocabulary segmentation - Box-prompted segmentation (from YOLO detections) - Fine-tuning on custom COCO-format datasets

Requirements: HuggingFace token (HF_TOKEN) for downloading gated weights.

Outputs: PNG overlays, COCO JSON annotations, YOLO segmentation .txt files.

Detectron2¶

Status: Supported Focus: COCO-format detection and instance segmentation Provider: Facebook Research

Model	Task	Backbone
Faster R-CNN R50-FPN 3x	Detection	ResNet-50 + FPN
RetinaNet R50-FPN 3x	Detection	ResNet-50 + FPN
Mask R-CNN R50-FPN 3x	Segmentation	ResNet-50 + FPN

Detectron2 training uses COCO-format annotations and integrates with ClearML for experiment tracking.

Model Selection Recommendations¶

Use RF-DETR When:¶

Need highest possible mAP (60.1 with 2XLarge)
Transformer-based real-time detection is desired
Working with high-resolution inputs
Server-side deployment with adequate compute

Use YOLO26 When:¶

Deploying on edge devices (CPU-only)
Need maximum inference speed
Memory is constrained
Working with IoT, robotics, aerial imagery

Use YOLOv12 When:¶

Researching attention mechanisms
Benchmarking new architectures
GPU resources available for training stability
Production deployments (not recommended)

Use YOLO11 When:¶

Need production-ready stability
Balance between speed and accuracy is needed
Training stability is critical
Deploying across varied environments

Use YOLOv10 or Earlier When:¶

Legacy system compatibility required
Specific pre-trained models unavailable in newer versions
Consistent behavior with existing pipelines needed

Use SAM 3.1 When:¶

Need open-vocabulary segmentation (no predefined classes)
Upgrading detection datasets to segmentation masks
Exploratory annotation and pseudo-label generation
Text-prompted concept segmentation

Use Detectron2 When:¶

Working with Faster R-CNN, RetinaNet, or Mask R-CNN architectures
Need COCO-format training with established baselines
Integration with the Detectron2 ecosystem

Supported Tasks¶

Task	Description	Supported Families
Detection	Identify and localize objects in images	RF-DETR, YOLO26, YOLOv12, YOLO11, YOLOv10, YOLOv9, YOLOv8, YOLOX, Detectron2
Segmentation	Detect objects and delineate their boundaries	RF-DETR, YOLO26, YOLOv12, YOLO11, YOLOv9, YOLOv8, SAM 3.1, Detectron2
Classification	Categorize images into predefined classes	YOLO26, YOLO11, YOLOv8
Pose Estimation	Detect and track keypoints on human bodies	YOLO26, YOLO11, YOLOv8
OBB	Detect rotated objects with higher precision	YOLO26, YOLO11, YOLOv8
Open-Vocab Seg	Segment objects by text concept (no predefined classes)	SAM 3.1

Architecture Evolution¶

YOLOv8 (2023)

YOLOv9 (Feature optimization)

YOLOv10 (NMS-free inference)

YOLO11 (Efficiency focus) ← Recommended for production

YOLO12 (Attention-centric) ← Research/benchmarking

YOLO26 (Edge optimization) ← Latest / Recommended for edge

Parallel families: - RF-DETR — transformer-based, highest mAP - SAM 3.1 — open-vocabulary segmentation - Detectron2 — Faster R-CNN / Mask R-CNN baselines

Performance Comparison Summary¶

Metric	RF-DETR	YOLO26	YOLO12	YOLO11	YOLOv10	YOLOv9
mAP (best variant)	60.1	57.5	55.2	54.7	54.4	55.6
CPU Speed (fastest)	—	Fastest	Slower	Fast	Moderate	Moderate
Params (smallest)	30.5M	2.4M	2.6M	2.6M	—	2.0M
Edge Optimized	No	Yes	No	Yes	Yes	Yes
Production Ready	Yes	Yes	No	Yes	Yes	Yes

Roboflow Upload Notes¶

YOLOmatic exposes a dedicated yolomatic-upload CLI for Roboflow uploads.
Upload a full checkpoint such as best.pt or last.pt, not generated artifacts like state_dict.pt.
RF-DETR .pth checkpoints deploy through RF-DETR's deploy_to_roboflow(...) flow and require a workspace, project ID, and project version.
YOLO26 uploads require a size-specific Roboflow model type such as yolo26n, yolo26s, yolo26m, yolo26l, or yolo26x.
Workspace defaults can be supplied with .env using ROBOFLOW_API_KEY, ROBOFLOW_WORKSPACE, and ROBOFLOW_PROJECT_IDS.
YOLO-NAS is deprecated in this build because SuperGradients conflicts with RF-DETR training dependencies.

Model Support Reference¶

Table of Contents¶

Quick Comparison¶

Benchmark Compatibility¶

RF-DETR¶

Detection Variants¶

Segmentation Variants¶

YOLO26¶

Key Features¶

Detection Performance (COCO val2017)¶

Segmentation Performance (COCO val2017)¶

Pose Performance (COCO Keypoints)¶

YOLOv12¶

Key Features¶

Detection Performance (COCO val2017)¶

Improvements vs Previous Versions¶

YOLO11¶

Key Features¶

Detection Performance (COCO val2017)¶

Segmentation Performance (COCO val2017)¶

Pose Performance (COCO Keypoints)¶

YOLOv10¶

YOLOv9¶

YOLOv8¶

Detection Performance¶

Segmentation Performance¶

Pose Performance (COCO Keypoints)¶

YOLOX¶

SAM 3.1¶

Detectron2¶

Model Selection Recommendations¶

Use RF-DETR When:¶

Use YOLO26 When:¶

Use YOLOv12 When:¶

Use YOLO11 When:¶

Use YOLOv10 or Earlier When:¶

Use SAM 3.1 When:¶

Use Detectron2 When:¶

Supported Tasks¶

Architecture Evolution¶

Performance Comparison Summary¶

Roboflow Upload Notes¶

References¶