Falcon Perception 300M
Falcon Perception 300M returns bounding boxes from an image and a natural-language query. This smaller variant supports detection only. It does not produce the pixel masks available from the full Perception model.
How it combines the prompt and image
Image patches and text enter the same transformer. The model generates coordinate and size tokens for each detected object. The implementation requires PyTorch 2.5 or newer for FlexAttention. Initial inference can include compilation work.Falcon Perception 300M model card
Use boxes when outlines are unnecessary
Practical guidance. Bounding boxes can be enough to crop a product, locate a field or pass a region to another model. Choose the full Perception release when the next step requires an exact object outline. Include cold-start time in a small-device comparison.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | tiiuae/Falcon-Perception-300M |
| License | Apache 2.0Falcon Perception 300M model card |
| Output | Normalized center coordinates and box sizesFalcon Perception 300M model card |
| Segmentation | Unsupported; requesting segmentation raises an errorFalcon Perception 300M model card |
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
Technology Innovation Institute (TII) · Falcon Perception
A 0.6B vision model that locates objects from text queries and returns pixel masks.
Launch date not confirmedDownloadable