Baidu releases Unlimited OCR 2 to read 100 pages at one time

Baidu released Unlimited OCR 2, an open model that reads text from images of pages. It has an MIT license, which is an open license, and reads a document of 100 pages at one time. Baidu tells that it gets 77.1 on MarathonDocBench, a test for long documents, but gave no other test result for the model.

How it works

The model reads a long document at one time. Baidu tells that its SR-SWA method keeps the cost of each decoding step almost flat. Decoding is the step in which the model writes its output.

The numbers

The one result is 77.1 on MarathonDocBench. The announcement gives no results for other models, no speed and no cost. At this time, a person cannot know if the model is good against the tools in use.

What is not known

It is not known how the model operates on your documents, for example scans or tables. It is also not known how much memory a document of 100 pages uses.

Insight: Big firms release the first steps of document tools free

Baidu uses an MIT license for a model that reads documents. Google released EmbeddingGemma 2, which finds text, images and audio, with an Apache 2.0 license. This reads as a move to hold the free first parts of document tools. The first parts that a company selects can possibly set the other parts. Builders can test the 2 parts together.