The French company LightOn has released LightOnOCR-3, a family of document models at three sizes: 0.8B, 1B and 4B parameters. Alongside transcription they now return bounding boxes with labels for document regions, descriptions of images, and numerical data read out of figures and charts, which the company pitches as a single-model alternative to a stack of separate document tools. The 1B keeps the previous architecture, while the 0.8B and 4B adopt the Qwen3.5 vision-language architecture.
The more useful part of the announcement is the part arguing against its own numbers. LightOn points out that the benchmarks it reports, ParseBench and olmOCR-Bench, score by edit distance, so two transcriptions a human would read as identical can score differently on formatting alone. At current accuracy levels, it says, a deterministic rewrite can shift a score noticeably without changing anything the model actually extracted.
It then says it deliberately did not tune its training data or output format to suit those conventions, which costs it points. The 4B lands 1.3 points behind a competing model listed at 35.1B parameters.
The throughput comparison is unusually legible too: the same 512 pages for every model, one H100 each, the same inference server and flags. No licence is named anywhere in the post, so that needs checking on the model repository before anyone builds on it.
Source: LightOn, 8 October 2026.