LightOn drops a tiny 1B OCR model that beats 9B giants
LightOn just open-sourced LightOnOCR-2-1B, an Apache 2.0 end-to-end vision-language model that turns PDFs and scans into clean text without multi-stage pipelines. Despite its tiny 1B footprint, it beats Chandra-9B on OlmOCR-Bench and runs 1.73x faster than DeepSeekOCR. The release also includes bounding-box variants and a massive 16M-page training dataset.
LightOnOCR-2-1B is an open-source vision-language model designed to turn messy PDFs, book scans, and complex documents into clean, structured text in a single step.
Instead of chaining multiple complex tools together, it uses a true end-to-end approach. It reads the page image and directly outputs naturally ordered text, handling tables, receipts, multi-column layouts, and raw LaTeX math without breaking a sweat.
Punching way above its weight
Even though it only has 1B parameters, LightOnOCR-2-1B delivers flagship-level transcription accuracy. On OlmOCR-Bench, it scored 83.2, outperforming the much larger Chandra-9B by more than 1.5 percentage points.
The real selling point here is sheer throughput. Running on a single NVIDIA H100 GPU, it rips through OlmOCR-Bench at 5.71 pages per second. In practice, that puts transcription costs at under a penny per 1,000 pages.
When it comes to raw inference speed, it leaves rivals behind. It is 1.7x faster than OlmOCR, 1.73x faster than DeepSeekOCR, 2x faster than PaddleOCR-VL-0.9B, and 3.3x faster than Chandra OCR.
What is in the release
LightOn released the entire family under the permissive Apache 2.0 license, meaning you can plug it straight into commercial pipelines.
You can pick the exact variant you need: the default LightOnOCR-2-1B for pure text transcription, a bbox variant that outputs bounding boxes for embedded figures and charts, or merged soup models that balance both.
They also open-sourced their training mix, including over 16 million annotated document pages and a 500k-sample bounding box dataset. You can try it out right away on their Hugging Face demo playground or grab the weights to self-host.
Why it matters
If you are building RAG pipelines, parsing scientific papers, or digitizing scanned archives, this model solves a massive real-world bottleneck.
You no longer need to burn API credits on giant proprietary vision models or maintain brittle multi-stage OCR pipelines. LightOn proves a lean 1B architecture is plenty capable of mastering complex document layouts at ultra-low inference costs.
Sources (1)
- huggingface.co(opens in a new tab)Originalhuggingface.co