TII Releases Falcon Perception: 0.6B Vision Grounding Model Outperforms SAM 3 on Compositional Tasks
Hugging Face Blog·high signal
Technology Innovation Institute released Falcon Perception, a 0.6B-parameter early-fusion Transformer for open-vocabulary grounding and segmentation from natural language prompts. It achieves 68.0 Macro-F1 on SA-Co benchmark (vs. 62.3 for SAM 3), with massive leads on OCR-guided identification (+13.4pts), spatial understanding (+21.9pts), and dense scenes (+14.2pts). Also ships Falcon OCR (0.3B, 80.3% on olmOCR) and PBench, a new diagnostic benchmark for perception capabilities.