LightOnOCR-2: a compact 1B-parameter document OCR model

LightOn released LightOnOCR-2-1B, a vision-language model for document OCR. It tops the OlmOCR-Bench while being 9x smaller than prior best models and processes 5.71 pages per second on H100 GPUs. The model is released under Apache 2.0.

• 1B parameters — 9x smaller than prior OlmOCR-Bench leaders
• Processes 5.71 pages per second on H100 GPUs
• Trained at up to 1540px on the longest edge
• Released under Apache 2.0 with dataset and benchmark