Know What's Happening in AI

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

The Technology Innovation Institute unveiled Falcon Perception, a 600‑million‑parameter dense transformer that fuses image patches and text tokens from the first layer, eliminating the traditional encoder‑decoder split. Using hybrid attention, 3D Rotary Positional Embeddings (GGROPE), and a Chain‑of‑Perception sequence, the model simultaneously learns visual representations and generates segmentation masks. Training involved multi‑teacher distillation and a 685‑gigatoken three‑stage pipeline, yielding strong gains on the new PBench benchmark, especially in spatial understanding. A 300‑million‑parameter variant, FalconOCR, matches larger OCR systems on document benchmarks.