YOLOv8
Ultralytics' 2023 YOLO, five sizes from n to x on COCO's 80 classes, AGPL-3.0 or a paid Enterprise license.
Conditional for size n: code AGPL-3.0-only, weights AGPL-3.0-only. Measured on our RTX 4090: 11.6 ms median, 514 MiB VRAM peak. We scored it 37.0 box AP on COCO val2017. PyTorch weights, 6 MB.
Conditional for size s: code AGPL-3.0-only, weights AGPL-3.0-only. Measured on our RTX 4090: 11.9 ms median, 582 MiB VRAM peak. We scored it 44.6 box AP on COCO val2017. PyTorch weights, 22 MB.
Conditional for size m: code AGPL-3.0-only, weights AGPL-3.0-only. Measured on our RTX 4090: 13.1 ms median, 710 MiB VRAM peak. We scored it 49.8 box AP on COCO val2017. PyTorch weights, 50 MB.
Conditional for size l: code AGPL-3.0-only, weights AGPL-3.0-only. Measured on our RTX 4090: 12.6 ms median, 846 MiB VRAM peak. We scored it 52.4 box AP on COCO val2017. PyTorch weights, 84 MB.
Conditional for size x: code AGPL-3.0-only, weights AGPL-3.0-only. Measured on our RTX 4090: 12.9 ms median, 1.0 GiB VRAM peak. We scored it 53.4 box AP on COCO val2017. PyTorch weights, 131 MB.
License
Conditional
Conditional
Conditional
Conditional
Conditional
- Code
- AGPL-3.0-onlychecked 3 Nov 2026
- Weights
- AGPL-3.0-onlyall sizes, checked 3 Nov 2026
AGPL-3.0: shipping it in a product means releasing your source, unless you buy an Ultralytics Enterprise license.
AGPL-3.0: shipping it in a product means releasing your source, unless you buy an Ultralytics Enterprise license.
AGPL-3.0: shipping it in a product means releasing your source, unless you buy an Ultralytics Enterprise license.
AGPL-3.0: shipping it in a product means releasing your source, unless you buy an Ultralytics Enterprise license.
AGPL-3.0: shipping it in a product means releasing your source, unless you buy an Ultralytics Enterprise license.
On our RTX 4090, measured
11.6ms median
p90 11.9 ms, fp16, batch 1, 640x448
- VRAM peak
- 514 MiB
- Status
- Runs as is
Environment, 3 Nov 2026
NVIDIA GeForce RTX 4090, driver 580.173.02, CUDA 13.0, torch 2.14.1+cu130, Python 3.12.9, ultralytics==8.4.189. Baseline 801 MiB. 10 warm-up + 100 timed. Scope: predict(): pre-processing, forward pass and post-processing; input file decoded beforehand.
Accuracy
37.0box APmeasured
COCO val2017, box AP, 5,000 images, fp16, pycocotools==2.0.11 · 2 Nov 2026
Paper 37.3 COCO mAP50-95Ultralytics YOLOv8 docs, detection performance table (mAPval 50-95, docs/en/models/yolov8.md at commit 4fb6485b)
11.9ms median
p90 12.2 ms, fp16, batch 1, 640x448
- VRAM peak
- 582 MiB
- Status
- Runs as is
Environment, 3 Nov 2026
NVIDIA GeForce RTX 4090, driver 580.173.02, CUDA 13.0, torch 2.14.1+cu130, Python 3.12.9, ultralytics==8.4.189. Baseline 801 MiB. 10 warm-up + 100 timed. Scope: predict(): pre-processing, forward pass and post-processing; input file decoded beforehand.
Accuracy
44.6box APmeasured
COCO val2017, box AP, 5,000 images, fp16, pycocotools==2.0.11 · 2 Nov 2026
Paper 44.9 COCO mAP50-95Ultralytics YOLOv8 docs, detection performance table (mAPval 50-95, docs/en/models/yolov8.md at commit 4fb6485b)
13.1ms median
p90 13.6 ms, fp16, batch 1, 640x448
- VRAM peak
- 710 MiB
- Status
- Runs as is
Environment, 3 Nov 2026
NVIDIA GeForce RTX 4090, driver 580.173.02, CUDA 13.0, torch 2.14.1+cu130, Python 3.12.9, ultralytics==8.4.189. Baseline 801 MiB. 10 warm-up + 100 timed. Scope: predict(): pre-processing, forward pass and post-processing; input file decoded beforehand.
Accuracy
49.8box APmeasured
COCO val2017, box AP, 5,000 images, fp16, pycocotools==2.0.11 · 2 Nov 2026
Paper 50.2 COCO mAP50-95Ultralytics YOLOv8 docs, detection performance table (mAPval 50-95, docs/en/models/yolov8.md at commit 4fb6485b)
12.6ms median
p90 12.9 ms, fp16, batch 1, 640x448
- VRAM peak
- 846 MiB
- Status
- Runs as is
Environment, 3 Nov 2026
NVIDIA GeForce RTX 4090, driver 580.173.02, CUDA 13.0, torch 2.14.1+cu130, Python 3.12.9, ultralytics==8.4.189. Baseline 801 MiB. 10 warm-up + 100 timed. Scope: predict(): pre-processing, forward pass and post-processing; input file decoded beforehand.
Accuracy
52.4box APmeasured
COCO val2017, box AP, 5,000 images, fp16, pycocotools==2.0.11 · 2 Nov 2026
Paper 52.9 COCO mAP50-95Ultralytics YOLOv8 docs, detection performance table (mAPval 50-95, docs/en/models/yolov8.md at commit 4fb6485b)
12.9ms median
p90 13.3 ms, fp16, batch 1, 640x448
- VRAM peak
- 1.0 GiB
- Status
- Runs as is
Environment, 3 Nov 2026
NVIDIA GeForce RTX 4090, driver 580.173.02, CUDA 13.0, torch 2.14.1+cu130, Python 3.12.9, ultralytics==8.4.189. Baseline 801 MiB. 10 warm-up + 100 timed. Scope: predict(): pre-processing, forward pass and post-processing; input file decoded beforehand.
Accuracy
53.4box APmeasured
COCO val2017, box AP, 5,000 images, fp16, pycocotools==2.0.11 · 2 Nov 2026
Paper 53.9 COCO mAP50-95Ultralytics YOLOv8 docs, detection performance table (mAPval 50-95, docs/en/models/yolov8.md at commit 4fb6485b)
Weights
Outputs on the shared test photos
Same photos and one score threshold for every model of this task. Failures are marked.
Sizes
| Size | Latency | VRAM | Accuracy | Weights | Runs | License |
|---|---|---|---|---|---|---|
| n | 11.6 ms640x448 | 514 MiB | 37.0 box AP, measured37.3 COCO mAP50-95 paper | 6 MB | As is | Conditional |
| s | 11.9 ms640x448 | 582 MiB | 44.6 box AP, measured44.9 COCO mAP50-95 paper | 22 MB | As is | Conditional |
| m | 13.1 ms640x448 | 710 MiB | 49.8 box AP, measured50.2 COCO mAP50-95 paper | 50 MB | As is | Conditional |
| l | 12.6 ms640x448 | 846 MiB | 52.4 box AP, measured52.9 COCO mAP50-95 paper | 84 MB | As is | Conditional |
| x | 12.9 ms640x448 | 1.0 GiB | 53.4 box AP, measured53.9 COCO mAP50-95 paper | 131 MB | As is | Conditional |