UAE’s TII Releases Falcon Perception Vision-Language Model Under Apache 2.0

ABU DHABI — The Technology Innovation Institute, the United Arab Emirates’ government-backed research center, has released Falcon Perception, a compact vision-language model that adds open-vocabulary object detection and instance segmentation to the Falcon open-source model family.

The 600-million-parameter model, published on Hugging Face under an Apache 2.0 license, uses an early-fusion architecture that processes image patches and text tokens together through a single dense Transformer backbone from the first layer. TII said the design allows users to identify and segment objects in images using natural language prompts without fixed label sets.

Falcon Perception introduces what TII calls Chain-of-Perception decoding, generating task tokens in a fixed sequence for coordinates, size and segmentation. A parallel mask decoding step then produces full-resolution binary masks, avoiding the slower autoregressive mask generation used by some competing approaches.

On TII’s internal SA-Co benchmark for open-vocabulary segmentation, Falcon Perception scored 68.0 macro F1, compared with 62.3 for Meta’s SAM 3. The institute also released PBench, a diagnostic dataset designed to test performance across spatial reasoning, OCR-guided disambiguation and dense crowded scenes.

The release includes a companion model, Falcon-OCR, optimized for optical character recognition tasks. Both models require PyTorch 2.5 or later and support torch.compile optimization for faster inference.

TII acknowledged limitations in the technical report published on arXiv. The model is more prone to false positives on hard negatives than detector-based architectures such as DETR, and OCR-driven prompts can struggle with small text or degraded scans. Dense scenes also require higher-resolution inputs for precise localization.

Falcon Perception has recorded more than 17,000 downloads in its first weeks on Hugging Face. The model’s code is available on GitHub.

The release marks TII’s latest effort to position the Falcon series as a competitive open-source alternative in a multimodal AI landscape increasingly dominated by proprietary systems from OpenAI, Google and Anthropic.