Skip to main content
Back

LightOn's LightOnOCR-2 Beats 9B Models at OCR While Running 3.3x Faster

LightOn's 1B end-to-end OCR model tops OlmOCR-Bench at 83.2, beats Chandra-9B while running nine times smaller and multiples faster.

  • LightOn released LightOnOCR-2-1B, a 1B end-to-end vision-language OCR model, Apache 2.0.
  • Scores 83.2 on OlmOCR-Bench, beating Chandra-9B by 1.5 points at 9x smaller.
  • Runs at 5.71 pages/sec on one H100, roughly $0.01 per 1,000 pages.
  • Family includes bbox variants, base checkpoints, and model-soup merges for fine-tuning.
  • Training data released: 16M-page mix plus 500k bbox annotations.
  • Supported natively in Transformers v5, vLLM, SGLang, and Docker Model Runner.

Document OCR has become one of the most competitive benchmarks in vision-language work, with new entries every few weeks from Allen AI, DeepSeek, Baidu, and a growing pile of pipeline-based systems. LightOn just released LightOnOCR-2-1B, a 1B-parameter vision-language model that takes a page image in and emits clean, ordered text out, with no layout detector or reading-order heuristic in the middle.

It achieves state-of-the-art performance on OlmOCR-Bench while running roughly nine times smaller and significantly faster than competing approaches. The full family ships under Apache 2.0, along with training datasets and a fine-tuning notebook.

What shipped

LightOnOCR-2 is a small family of checkpoints, so you can pick a tradeoff rather than retrofit one model to every job:

  • LightOnOCR-2-1B: the default, best pure transcription quality
  • LightOnOCR-2-1B-bbox: adds bounding-box outputs for embedded figures and images
  • LightOnOCR-2-1B-ocr-soup and -bbox-soup: model-soup merges that balance both objectives
  • LightOnOCR-2-1B-base and -bbox-base: pre-RLVR checkpoints intended for fine-tuning or reproducing the post-training recipe

LightOn also released two open annotation datasets used during training: LightOnOCR-mix-0126, with more than 16M annotated document pages, and LightOnOCR-bbox-mix-0126, with close to 500k annotations including bounding boxes for figures and images.