About yolo-models
A reference skill for choosing Ultralytics vision models by task, size, and deployment constraints. It helps users select model families and configure custom architectures.
Key features
- 01Compare model families and sizes
- 02Match suffixes to vision tasks
- 03Select open-vocabulary or promptable models
- 04Load and customize architecture YAMLs
Use cases
- 01Choose a model for a new vision task
- 02Select a model size for latency constraints
- 03Decide between detection and segmentation approaches
Install
Copy the command and run it in a terminal.
npx skills add https:/After installing, type/yolo-modelsin Claude Code to confirm it loaded.
Bundled filesagents/openai.yamlweights-catalog.md
Details
From the project READMEChoosing an Ultralytics model
Default recommendation: YOLO26, pretrained. Latest generation, with an optional NMS-free head (nms=False) for the fastest CPU inference and simplest deployment. Use YOLO11/YOLOv8 only to match an existing codebase or a deployment target that doesn't support YOLO26 yet. Most official weights auto-download on first use; sam3.pt and sam3.1_multiplex.pt require manual access and download.
Choose in Platform
For the quickest no-code start, click New Model in a project. Its Official tab lists YOLO26 (recommended), YOLO11, YOLOv8, and YOLOv5 models whose task can train the selected dataset; My Models lists your completed or uploaded checkpoints for further fine-tuning. To browse the official models first, open the @ultralytics projects under Projects in Platform Explore.
Use a Platform model page to inspect metrics, test it in Predict, export it, deploy it, clone a public model into your workspace, or download its .pt weights for the Python/CLI workflows below. See Platform Models and Explore.
Model = family + size + task suffix
yolo26 + n/s/m/l/x + task suffix → yolo26s-seg.pt
| Size | COCO mAP50-95 | Params | T4 TensorRT | Pick for |
|---|---|---|---|---|
| n | 40.9 | 2.4M | ~1.7 ms | edge/mobile, CPU realtime, first prototype |
| s | 48.6 | 9.5M | ~2.5 ms | balanced default for most projects |
| m | 53.1 | 20.4M | ~4.7 ms | GPU server, accuracy matters |
| l | 55.0 | 24.8M | ~6.2 ms | accuracy-critical, ample GPU |
| x | 57.5 | 55.7M | ~11.8 ms | max accuracy, offline/batch |
Strategy: prototype on n to validate the pipeline cheaply, then scale up until accuracy stops paying for the latency. A bigger model never fixes bad labels.
| Suffix | Task | Output |
|---|---|---|
| (none) | detect | boxes |
-seg | instance segmentation | polygons + boxes |
-sem | semantic segmentation (YOLO26+) | per-pixel class mask |
-depth | monocular depth (YOLO26+) | depth map |
-cls | classification | class probabilities |
-pose | pose/keypoints | keypoints + boxes |
-obb | oriented boxes | rotated boxes |
Notes on the newer tasks:
- semantic (
-sem): dataset uses PNG masks viamasks_dir(defaultmasks/) or polygon labels; metric is mIoU. - depth (
-depth): targets are scaled uint16 PNG maps (preferred) or floating-point.npymaps in meters; metric is delta1. Depth models also exposemodel.calibrate(data=...), which fits a metric-scale correction; persist it withmodel.save(...).
Family cheat sheet
| Family | Class | When |
|---|---|---|
| YOLO26 / YOLO11 / YOLO12 / YOLOv8–v10 | YOLO("yolo26n.pt") | standard closed-set tasks; default choice |
| YOLO-World | YOLOWorld("yolov8s-worldv2.pt") | zero-shot detection of arbitrary text classes; model.set_classes(["person", "helmet"]); prefer -worldv2 (original -world weights cannot export) |
| YOLOE | YOLOE("yoloe-26s-seg.pt") | open-vocabulary detect+segment via text or visual prompts; set_classes(["person", "helmet"]), visual prompts via predict(..., visual_prompts={"bboxes": ..., "cls": ...}, predictor=YOLOEVPSegPredictor) (the default predictor returns boxes only); -pf variants are prompt-free |
| SAM / SAM2 / SAM3 / MobileSAM | SAM("sam_b.pt") | promptable segmentation: predict(source, bboxes=... / points=... / labels=...); SAM2/SAM3 add video predictors; SAM3 adds text/exemplar concept segmentation (SAM3SemanticPredictor); inference only |
| FastSAM | FastSAM("FastSAM-s.pt") | CNN-based segment-anything, much faster than SAM; no training |
| RT-DETR | RTDETR("rtdetr-l.pt") | transformer detector, strong accuracy on GPU |
| YOLO-NAS | NAS("yolo_nas_s.pt") | no training (val/predict/export only) |
All classes share the same Model API (train/val/predict/track/export/...) — everything in the other yolo-* skills applies to them, with the exceptions noted above.
Open-vocabulary decision: need arbitrary classes at inference with no training → YOLO-World (detect) or YOLOE (detect+segment, also visual prompts). Need pixel-precise masks from clicks/boxes → SAM family. Need a trained model for a fixed class list → plain YOLO26 fine-tune (faster and more accurate on that closed set).
Architecture YAMLs (custom models)
ultralytics/cfg/models/ ships editable architecture definitions (yolo26.yaml, yolo11.yaml, yolov8.yaml, and variants such as -p2 (small objects), -p6 (large imgsz), and -ghost). Loading YOLO("yolo26n.yaml") builds from scratch — scale is picked from the letter in the stem. To customize the architecture but keep pretrained weights where layers match:
model = YOLO("yolo26n.yaml").load("yolo26n.pt") # transfer matching weights
Only go here for research/unusual constraints; for normal work fine-tune the stock .pt.
Related pages
weights-catalog.md(this folder) — read for package-known weight patterns and specialized official assets. Do not guess weight names.
Verify against the installed version
Model availability moves fast. This prints the installed package's known fast-path set; read weights-catalog.md before treating an unlisted official asset as invalid:
python -c "from ultralytics.utils.downloads import GITHUB_ASSETS_NAMES; print(*sorted(GITHUB_ASSETS_NAMES), sep='\\n')"
If a weight 404s or a class import fails, check yolo checks and trust the installed-version error.