TII Releases Falcon Perception: A 0.6B-Parameter Early-Fusion Transformer for Open-Vocabulary Grounding and Segmentation from Natural Language Prompts

TII released Falcon Perception, a Transformer that feeds image patches and text tokens into one unified stack.

It performs open-vocabulary grounding and segmentation from natural-language prompts.

The model uses hybrid attention, 3D rotary positional embeddings, and a chain-of-perception output format.

It was trained with multi-teacher distillation in a three-stage pipeline totaling about 685 billion tokens.

The release suggests a single early-fusion Transformer can compete with modular vision-plus-decoder designs.