Comparison · 7 YOLO models
YOLO model comparison
Every major YOLO model side by side: published COCO accuracy, parameter counts, architecture (anchor-free, NMS-free / end-to-end), the tasks each supports, and licensing. All figures are the authors' own published numbers, sourced below.
- Best all-round defaultYOLO11
- Highest published accuracyYOLO26
- NMS-free / end-to-end edgeYOLO26 / YOLOv10
- Widest ecosystem & tutorialsYOLOv8
53.1 COCO val2017 mAP50-95 @ 640px on its YOLO26m — the highest published here.
Anchor-free, five tasks, the largest community and tutorials — the safe production pick today.
Output final detections with no NMS step — simpler export and lower latency for edge and CPU.
Accuracy at a glance
COCO val2017 mAP50-95 @ 640px, higher is better. The number on the right is parameter count (smaller is lighter). These are the authors' published figures — indicative of tier, not a re-run benchmark.
Smallest variant (nano tier)
Medium variant
Full specification matrix
Architecture, tasks, sizes and license for all 7 models. Tracking (BoT-SORT / ByteTrack) is an Ultralytics ecosystem feature available to every integrated model.
| Model | Year | Anchor-free | NMS-free | Tasks | Nano mAP | Medium mAP | License |
|---|---|---|---|---|---|---|---|
| YOLO26 | 2026 | Yes | Yes | 5 (Detect, Segment, Classify, Pose, OBB) | 40.9 | 53.1 | AGPL-3.0 |
| YOLO11 | 2024 | Yes | No | 5 (Detect, Segment, Classify, Pose, OBB) | 39.5 | 51.5 | AGPL-3.0 |
| YOLOv10 | 2024 | Yes | Yes | 1 (Detect) | 39.5 | 51.3 | AGPL-3.0 |
| YOLOv9 | 2024 | Yes | No | 2 (Detect, Segment) | 38.3 | 51.4 | GPL-3.0 (Ultralytics build: AGPL-3.0) |
| YOLOv8 | 2023 | Yes | No | 5 (Detect, Segment, Classify, Pose, OBB) | 37.3 | 50.2 | AGPL-3.0 |
| YOLOv7 | 2022 | No | No | 2 (Detect, Pose) | 38.7 | — | GPL-3.0 |
| YOLOv5 | 2020 | No | No | 3 (Detect, Segment, Classify) | 28.0 | 45.4 | AGPL-3.0 |
Every model in depth
YOLO26
2026 · UltralyticsEnd-to-end and NMS-free by default, tuned for the edge.
YOLO26 is Ultralytics' end-to-end detector: it drops NMS and Distribution Focal Loss (DFL) so the model outputs final predictions directly, which simplifies export and speeds up CPU and edge inference. It keeps the full multi-task head (detect, segment, pose, classify, OBB) and posts the highest published COCO accuracy of any model here at a comparable size.
- Key idea
- NMS-free + DFL-free end-to-end head, MuSGD training
- Best for
- New builds that want top accuracy and clean edge/CPU deployment.
- Sizes
- n, s, m, l, x
- + Highest published COCO mAP at each size here
- + NMS-free end-to-end — simpler, faster export
- + DFL-free head is friendlier to edge runtimes
- − Newest model — smallest community and fewest tutorials
- − Published figures may still be updated as it settles
YOLO11
2024 · UltralyticsThe all-round Ultralytics default across five vision tasks.
YOLO11 is the mainstream Ultralytics model: anchor-free, well-documented, and supporting all five vision tasks. Its C3k2 blocks and C2PSA attention let YOLO11m match YOLOv8m accuracy with about 22% fewer parameters, which makes it the safe, best-supported default for most production work today.
- Key idea
- C3k2 blocks + C2PSA spatial attention
- Best for
- The best-documented general-purpose default for production.
- Sizes
- n, s, m, l, x
- + Excellent accuracy-per-parameter
- + All five tasks, huge docs & community
- + Battle-tested across production deployments
- − Still needs NMS (not end-to-end)
- − YOLO26 now edges it on published accuracy
YOLOv10
2024 · Tsinghua University (THU-MIG)The first popular NMS-free, end-to-end YOLO.
YOLOv10 (Tsinghua University, NeurIPS 2024) introduced consistent dual assignments — a one-to-many head for training and a one-to-one head for NMS-free inference — making it the first widely used end-to-end YOLO. It's detection-only, but its efficiency-first design gives strong accuracy at low parameter counts and low latency.
- Key idea
- Consistent dual assignments for NMS-free inference
- Best for
- Low-latency detection where end-to-end export matters.
- Sizes
- n, s, m, b, l, x
- + NMS-free, end-to-end inference
- + Very low parameter count for its accuracy
- + Strong latency on constrained hardware
- − Detection only — no seg/pose/classify/OBB
- − Smaller ecosystem than the Ultralytics-native models
YOLOv9
2024 · Academia Sinica (Wang et al.)Fixes deep-network information loss with PGI + GELAN.
YOLOv9 (Academia Sinica, ECCV 2024) tackles the information bottleneck in deep networks with Programmable Gradient Information (PGI) and the Generalized ELAN (GELAN) architecture. It delivers high accuracy at low parameter counts and is integrated into Ultralytics for detection and segmentation.
- Key idea
- Programmable Gradient Information (PGI) + GELAN
- Best for
- Squeezing high accuracy out of a small parameter budget.
- Sizes
- t, s, m, c, e
- + Excellent accuracy for very few parameters
- + PGI improves training of deep models
- + Detection + segmentation via Ultralytics
- − No pose / classify / OBB
- − GPL-3.0 original repo is more restrictive
YOLOv8
2023 · UltralyticsThe workhorse that made anchor-free multi-task YOLO mainstream.
YOLOv8 was the model that popularized anchor-free, multi-task YOLO across the industry. Its C2f backbone and decoupled head cover all five tasks, and years of tutorials, integrations and deployments mean it remains the most widely referenced YOLO in the wild — even as YOLO11 and YOLO26 surpass it on accuracy-per-parameter.
- Key idea
- C2f blocks + anchor-free decoupled head
- Best for
- Maximum ecosystem support and third-party integrations.
- Sizes
- n, s, m, l, x
- + Vast community, tutorials and integrations
- + All five tasks, extremely well proven
- + Stable, predictable behaviour
- − Beaten on accuracy-per-parameter by YOLO11/YOLO26
- − Heavier than newer models at the same accuracy
YOLOv7
2022 · Academia Sinica (Wang et al.)The 2022 anchor-based accuracy leader.
YOLOv7 (Academia Sinica) set the real-time detection state of the art in 2022 with trainable bag-of-freebies, model re-parameterization and E-ELAN. It's anchor-based and NMS-based, uses its own variant scheme rather than n/s/m/l/x, and its GPL-3.0 license and separate task branches make it less turnkey than the Ultralytics line.
- Key idea
- Trainable bag-of-freebies + E-ELAN + re-parameterization
- Best for
- Reproducing 2022-era research baselines.
- Sizes
- tiny, base, X, W6/E6/D6/E6E
- + Strong accuracy for its era
- + Well-cited research baseline
- − Anchor-based, NMS-based — older design
- − No clean size ladder; segmentation/pose live in side branches
- − GPL-3.0; not natively trainable in Ultralytics
YOLOv5
2020 · UltralyticsThe classic that made YOLO easy to ship in PyTorch.
YOLOv5 made YOLO practical for everyone: pure PyTorch, easy training and export, and a huge deployment footprint that persists today. As originally released it is anchor-based and NMS-based; Ultralytics later shipped an anchor-free retrain (YOLOv5u) that adopts the YOLOv8 head and raises accuracy. The numbers here are the original anchor-based release.
- Key idea
- First fully PyTorch YOLO; CSPDarknet + PANet
- Best for
- Legacy pipelines and the widest deployment footprint.
- Sizes
- n, s, m, l, x (+ P6)
- + Enormous install base and tooling
- + Simple to train, export and deploy
- + Anchor-free YOLOv5u retrain available
- − Original release is anchor-based and lower accuracy
- − Superseded by YOLOv8/YOLO11 for new work
How to choose a YOLO model
Starting a new production project
Pick YOLO11: the best-documented, best-supported model with all five tasks and strong accuracy-per-parameter.
You want the newest and most accurate
Pick YOLO26: highest published COCO mAP here, NMS-free and DFL-free for clean edge and CPU export.
Deploying to edge / CPU with tight latency
Pick YOLO26 or YOLOv10: both are NMS-free end-to-end, so there's no post-processing step to slow inference.
You rely on third-party tools & tutorials
Pick YOLOv8: the widest ecosystem, integrations and community answers, still very capable.
Maximum accuracy from tiny models
Look at YOLOv9: PGI + GELAN give strong mAP at very low parameter counts.
Maintaining a legacy pipeline
Stay on YOLOv5 (or move to YOLOv5u) if migration cost outweighs the accuracy gains of newer models.
Compare any two models
Jump straight to any pairwise matchup. Each page compares the two models on accuracy, size, architecture, tasks and license.
- YOLO26 vs YOLO11
- YOLO26 vs YOLOv10
- YOLO26 vs YOLOv9
- YOLO26 vs YOLOv8
- YOLO26 vs YOLOv7
- YOLO26 vs YOLOv5
- YOLO11 vs YOLOv10
- YOLO11 vs YOLOv9
- YOLO11 vs YOLOv8
- YOLO11 vs YOLOv7
- YOLO11 vs YOLOv5
- YOLOv10 vs YOLOv9
- YOLOv10 vs YOLOv8
- YOLOv10 vs YOLOv7
- YOLOv10 vs YOLOv5
- YOLOv9 vs YOLOv8
- YOLOv9 vs YOLOv7
- YOLOv9 vs YOLOv5
- YOLOv8 vs YOLOv7
- YOLOv8 vs YOLOv5
- YOLOv7 vs YOLOv5
How this comparison is built
- Published numbers only. Every mAP and parameter figure is the model authors' own published value (COCO val2017 mAP50-95 @ 640px), linked from each model's card — nothing here is measured on this site.
- Indicative, not a controlled benchmark. The families use different size ladders (YOLOv9's smallest is
t, YOLOv7 has no standard medium, YOLOv5's figures are the original anchor-based release), so cross-family mAP shows the tier, not a like-for-like race. Benchmark your shortlist on your own data. - Architecture facts. Anchor-free, NMS-free, tasks and license reflect each model as officially released.
- Want a real speed test? See the YOLO tracker benchmarks for measured FPS across 10 GPUs, or book a consultation to benchmark models on your footage.
From the blog
Tutorials, code, and notes on computer vision, deep learning, and applied AI.
YOLO26 vs YOLO11: Real-time ONNX FPS benchmark in Python
Build one small, reusable class that runs Ultralytics YOLO26 and Ultralytics YOLO11 as ONNX models, draws clean detections, and overlays live FPS and latency so you can compare their real-time speed on the exact same footage.
Ultralytics object trackers comparison: ByteTrack, BoT-SORT & More
How do the six Ultralytics trackers actually behave on the same footage? A look at BoT-SORT, ByteTrack, OC-SORT, Deep OC-SORT, FastTrack, and TrackTrack, their internals, trade-offs, and side-by-side results on ID switches, ID stability, and FPS.
Object tracking and trajectory forecasting with YOLO26 and ByteTrack
Detect, track, and predict the future path of people and vehicles using Ultralytics YOLO26, ByteTrack, and a lightweight velocity-based forecasting model.
Real time bird detection and tracking using YOLO11
Learn how to detect and track birds using Ultralytics YOLO11 for real-time monitoring and ecological research through computer vision.
How to count people in zones with YOLO26 and OpenCV
A practical walkthrough of a compact Python script that detects, tracks, and counts people inside polygon zones using Ultralytics YOLO26 and OpenCV.
Supermarket items segmentation and counting with YOLO11
Discover how to achieve advanced items segmentation in supermarkets with Ultralytics YOLO11 for efficient object detection and analysis.
How to generate accurate segmentation masks using object detection and Meta SAM2
Learn how to generate segmentation masks with object detection and SAM2 models for advanced image processing tasks.
YOLO26 vs YOLO11: Real-time ONNX FPS benchmark in Python
Build one small, reusable class that runs Ultralytics YOLO26 and Ultralytics YOLO11 as ONNX models, draws clean detections, and overlays live FPS and latency so you can compare their real-time speed on the exact same footage.
Ultralytics object trackers comparison: ByteTrack, BoT-SORT & More
How do the six Ultralytics trackers actually behave on the same footage? A look at BoT-SORT, ByteTrack, OC-SORT, Deep OC-SORT, FastTrack, and TrackTrack, their internals, trade-offs, and side-by-side results on ID switches, ID stability, and FPS.
Object tracking and trajectory forecasting with YOLO26 and ByteTrack
Detect, track, and predict the future path of people and vehicles using Ultralytics YOLO26, ByteTrack, and a lightweight velocity-based forecasting model.
Real time bird detection and tracking using YOLO11
Learn how to detect and track birds using Ultralytics YOLO11 for real-time monitoring and ecological research through computer vision.
How to count people in zones with YOLO26 and OpenCV
A practical walkthrough of a compact Python script that detects, tracks, and counts people inside polygon zones using Ultralytics YOLO26 and OpenCV.
Supermarket items segmentation and counting with YOLO11
Discover how to achieve advanced items segmentation in supermarkets with Ultralytics YOLO11 for efficient object detection and analysis.
How to generate accurate segmentation masks using object detection and Meta SAM2
Learn how to generate segmentation masks with object detection and SAM2 models for advanced image processing tasks.
Frequently asked questions
- Which YOLO model is the most accurate?
- By published COCO mAP at a comparable size, YOLO26 leads (53.1 mAP for the medium model), just ahead of YOLO11 (51.5) and YOLOv9 (51.4). All numbers here are the authors' own published figures at 640px, not a re-run benchmark, so treat small gaps as ties and test on your own data.
- What is the difference between anchor-free and NMS-free?
- Anchor-free means the model predicts boxes directly instead of refining pre-set anchor boxes (YOLOv8 onward). NMS-free (end-to-end) goes further: the model outputs final detections with no Non-Maximum-Suppression post-processing step, which simplifies export and cuts latency. YOLOv10 and YOLO26 are NMS-free; YOLOv8, YOLOv9 and YOLO11 still use NMS.
- Should I use YOLO11 or YOLO26?
- YOLO11 is the safest default today: it has the largest community, the most tutorials, and covers all five tasks. YOLO26 is newer, posts higher published accuracy, and is NMS-free and DFL-free for cleaner edge and CPU deployment. Choose YOLO11 for maximum support, YOLO26 for a fresh build that wants the newest architecture.
- Are these mAP numbers directly comparable across models?
- They are indicative, not a controlled benchmark. Every figure is the authors' own published COCO val2017 mAP at 640px, but the families don't share the same size ladder (YOLOv9's smallest is 't', YOLOv7 has no 'medium', YOLOv5's numbers are the original anchor-based release). Use them to see the general tier, then benchmark the shortlist on your own hardware and data.
- Which YOLO license can I use commercially?
- Most models here are AGPL-3.0 (Ultralytics YOLOv5, YOLOv8, YOLOv10, YOLO11, YOLO26), which requires open-sourcing derivative network-served applications unless you buy an Ultralytics Enterprise license. YOLOv7 and the original YOLOv9 repo are GPL-3.0. Always check the license against your deployment before shipping.