Hugging Face releases Falcon Perception and OCR
hugging face released falcon perception, a unified transformer architecture for vision tasks, and falcon ocr, a 0.3b-parameter model. falcon ocr scored 80.3 on olmocr and 88.6 on omnidocbench, achieving the highest throughput of any open source ocr model. the architecture uses early fusion and chain-of-perception to handle dense predictions without modular pipelines.
Hugging Face released Falcon Perception, an open-source early-fusion Transformer architecture for perception tasks, alongside Falcon OCR. Falcon OCR is a 0.3B-parameter model that achieved scores of 80.3 on the olmOCR benchmark and 88.6 on OmniDocBench. Hugging Face states the model delivers the highest throughput of any open-source OCR model currently available.
The Falcon Perception architecture abandons modular pipelines in favor of a single autoregressive backbone. It employs a hybrid attention mask where image tokens attend bidirectionally while text tokens attend causally. The system uses a 'Chain-of-Perception' method to decompose dense predictions into three sequential steps: coordinates, size, and segmentation masks.