Fast, efficient, state-of-the-art document understanding in a single model

Traditional OCR pipelines combine several specialized components, combining models used for layout detection, transcription, additional logic for tables and charts, and custom processing to reconnect every extracted element to its position on the page… The systems can perform well, however they are difficult to build, maintain and scale.

LightOnOCR-3 brings us closer to our long standing goal of having a single model as a replacement for complete OCR pipelines. Additionally to improved transcription quality, it is trained to return a grounded representation of the page that provides layout structure, image descriptions as well as chart data extraction.

No more pipelines, a single model is all you need to build highly performant workflows for search, extraction or RAG applications.

LightOnOCR-3 delivers leading results across aggregated benchmarks for text recognition and the interpretation of visual information, they are additionally among the fastest on the market, processing documents twice as fast as LightOnOCR-2. They are also the cheapest among leading OCR models. When run on an organization’s own infrastructure, processing costs can fall below one cent per thousand pages, depending on hardware, usage levels and workload.

In self-hosted deployments, processing costs can fall below one cent per thousand pages, depending on hardware, utilization and workload, offering a low-cost high performance alternative to commercial OCR services such as Mistral OCR.

The models can be used as before in a transcription-only mode, where all textual elements of the page are outputted, if you’re currently using our previous models switching to v3 requires no change as it’s the default behavior when calling it with an empty prompt.

The new usage mode is to call the models with the grounding prompt. It will then output all the new vision features alongside transcribed text.

LightOnOCR 3 can identify document elements and return their bounding-box coordinates, preserving where paragraphs, tables, images and other blocks appear on the original page. Each block of content is now enriched depending on the nature of the block:

• Textual content blocks contain the transcribed content of paragraphs, titles and other text elements.
• Image blocks pair a bounding box with a short description, making visual content accessible to retrieval and question-answering pipelines.
• Chart blocks contain an HTML table of data points extracted from the figure, transforming visual information into structured data for reporting, analysis and retrieval.

The structured output unlocks several common workloads:

• RAG and enterprise search: Create cleaner, layout-aware chunks and retain the location of supporting content for grounded answers and citations.
• Document extraction: Recover text, tables, charts and other elements from complex business documents without maintaining separate parsers for each content type.
• Knowledge-base ingestion: Convert heterogeneous documents into consistent, typed content before indexing.
• Document agents: Give agents both the content of a document and the structural context needed to act on forms, reports and records.
• Scientific and technical document processing: Preserve equations, tables, figures and chart data that basic OCR pipelines often flatten or lose.