Falcon Perception, OCR land open-source release

This looks like a credible open-source bet on simpler multimodal architecture: smaller models, cleaner internals, and useful benchmark gains rather than another bloated pipeline stack.

• –Falcon Perception uses one shared Transformer for image patches and text, which is easier to reason about and potentially easier to scale than stitched-together vision systems
• –Falcon OCR is the more immediately practical release for developers, with strong claims on table extraction, multi-column docs, handwriting, and throughput
• –The release matters beyond raw scores because the ecosystem story is real: Hugging Face hosting is live, and local inference work like llama.cpp support is already underway
• –The benchmark claims are promising, but they still need independent validation before anyone treats them as a new default for OCR or segmentation