LightOn Releases OCR-2, a 1B Document AI Model
The new vision model from the Paris-based AI lab uses Mistral architecture to extract text and structure from complex documents like PDFs and forms.
Parisian AI company LightOn has released LightOnOCR-2, a new 1-billion-parameter vision language model specialized in document understanding. The model is designed to perform optical character recognition (OCR) on complex documents, extracting not just text but also structural information.
Unlike simple OCR tools, LightOnOCR-2 is built to parse challenging layouts like tables, forms, and multi-column PDFs. This makes it suitable for enterprise automation tasks such as processing invoices or digitizing records, a domain often dominated by proprietary, API-gated services.
A Mistral-based Architecture
The model features a Transformer-based encoder-decoder architecture. In an interesting design choice, its decoder was initialized using a subset of weights from Mistral-7B-v0.1, allowing it to leverage the powerful language capabilities of the popular open model while maintaining a much smaller, more efficient footprint.
The complete model weights and code are available on the Hugging Face Hub for developers to download and use. It's released under a custom LightOnAI-OpenRAIL-M license, which permits commercial use but includes some use-case restrictions common to Responsible AI licenses.
Sources
- Visit
lightonai/LightOnOCR-2-1B
Hugging Face
More from LightOn
All LightOn releases →
LightOn ships a 4B OCR model with visual grounding
LightOnOCR-3-4B pairs document transcription with the ability to point back to where text sits on the page.
More in Vision-Language
All Vision-Language →
Mistral Large 4 Arrives as a 1T-Parameter MoE
The new flagship pairs a trillion total parameters with 52B active per token and adds vision alongside multilingual text.
Baidu releases Unlimited-OCR 2 under MIT license
The multilingual vision-language model targets document text recognition and ships with a permissive license for commercial use.
Liquid AI's LFM3-Omni-2B brings any-to-any AI to the edge
The 2-billion-parameter model combines vision, audio and speech-to-speech in a package small enough to run on device.
0 comments
No comments yet. Be the first to weigh in.