Falcon Perception
Falcon Perception uses a text query to locate objects in an image. The 0.6B model returns matching instances with pixel masks. It is an open-vocabulary grounding model, not a general chat assistant.
A revision for crowded images
The August revision uses GRPO with a reward that matches predicted and reference objects. The card reports better recall in dense scenes, including images with up to 500 instances. The architecture and tokenizer remain the same. Small text and false detections remain documented limitations.Falcon Perception model card and RL revision
Ask for the objects your workflow needs
Practical guidance. Use precise object descriptions and include images where the requested object is absent. A mask model must both find the correct item and avoid selecting a similar one. Record the checkpoint revision so later downloads do not change the comparison.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | tiiuae/Falcon-Perception |
| License | Apache 2.0Falcon Perception model card and RL revision |
| Output | Object coordinates, sizes and segmentation masksFalcon Perception model card and RL revision |
| Additional release | RL checkpoint at revision 19-08-2026Falcon Perception model card and RL revision |
Related models
- Falcon OCR 1.5
Technology Innovation Institute (TII) · Falcon OCR
Falcon OCR reads document images with a 270M parameter model.
Launch date not confirmedDownloadable - Falcon Perception 300M
Technology Innovation Institute (TII) · Falcon Perception
A 300M vision model that returns object bounding boxes from images and text queries.
Launch date not confirmedDownloadable