LightOn

As is almost tradition at this point, the French company LightOn AI has released another multi-vector model based on ModernBERT with SOTA performance.

A very capable OCR model, reaching SOTA performance even against models which are up to 9 times its size.

The increase of open OCR model releases is really commendable. The French LightOn is throwing their version into the ring, licensed under Apache 2.0.

A SOTA retrieval model released under a non-commercial license. However, there is also code to re-generate the data, allowing the training of a commercially viable version.

A multi-vector embedding model based on ModernBERT.