Supported Models
All models you can deploy with Roboflow.
With Roboflow, you can deploy object detection, segmentation, classification, keypoint, Action Recognition, and multimodal models. Models you have trained are referenced by a model ID.
These models can be deployed with Workflows using:
Self-hosted (Roboflow Inference)
The table below shows common models and their support for training, model weights upload and model weights export/download, and Serverless Cloud API deployment. For models that support weights export, the file formats available (ex: ONNX, PyTorch, CoreML) vary by architecture: see the export format table.
The table describes models that can be deployed with the Serverless Cloud API.
All models can be used with Batch Processing.
We recommend using Dedicated Deployments or Batch Processing with a GPU configured for optimal performance when using multimodal models.
Object Detection
Instance Segmentation
Action Recognition
Train Cosmos 3 Edge on labeled video segments and run the trained model through the Serverless Cloud API to get a timeline of actions.
You can also train V-JEPA 2.1 (Beta) on a paid plan. Pick it in the training picker. It supports Workflows, model evaluation, and weight download, but not auto-labeling. You can stop a V-JEPA 2.1 training early and continue it later from its last checkpoint.
Classification
Keypoint Detection
Semantic Segmentation
Multimodal
OCR
Other
Models Supported in Workflows
You can run all of the above models in Workflows, as well as other models like Segment Anything 2, CLIP, OpenAI's GPT models, and more. See a full list of models you can run in a Workflow.
Foundation Models
Foundation models are trained on very large, general datasets, so they work out of the box without you training anything. OpenAI's CLIP, for example, was trained on over 400 million image-text pairs, which lets it classify and compare images across a wide range of domains.
Roboflow supports foundation models for segmentation, zero-shot detection, embeddings, OCR, depth estimation, and visual question answering. Every model listed on this page is callable over HTTP, so you do not have to set up or configure each one yourself.
Common uses:
Zero-shot detection and segmentation. SAM3, Grounding DINO, YOLO-World, and OWLv2 find objects from a text prompt or a single example, with no training data.
Automated labeling. Use a zero-shot model to label a dataset, then train a smaller, faster model that is fine-tuned to your use case.
Embeddings. CLIP and Perception Encoder turn images and text into vectors for clustering, deduplication, content moderation, and dataset search.
Reading and describing images. Florence 2, Qwen3-VL, Moondream2, and the OCR models answer questions about images and extract text.
Pretrained Models
Roboflow ships COCO-pretrained checkpoints under short aliases such as rfdetr-small and yolov8n-640. You can pass an alias anywhere a model ID is accepted, with no training and no project setup. See Pretrained Model Aliases for the full list.
Running Your Own Models
To run a model you trained (or uploaded weights for), you need its model ID. Open your Roboflow dashboard, select the model, and click "Deploy" in the sidebar to find the ID, which has the form {workspace}/{project}/{version}.
Then pass it as model_id:
Load the results into supervision with sv.Detections.from_inference(results) to visualize them.
Community Models
Roboflow Universe hosts more than 50,000 models published by other Roboflow users, covering everything from defect detection to sports analytics and wildlife identification. Universe model IDs work exactly like your own: see Find a Model on Universe.
Licensing
Models carry the licenses of their upstream projects, shown in the License column. RF-DETR is Apache 2.0, so you can use it commercially, including self-hosted, with no license fee (RF-DETR XL and 2XL detection use PML 1.0). YOLOv5, YOLOv8, YOLOv10, YOLO11, YOLOv12, YOLO26, and YOLO-World are AGPL-3.0: deploying them on Roboflow cloud includes a commercial license for every user, and commercial self-hosting is an Enterprise add-on. See Model Licenses.
Last updated
Was this helpful?
