Four open OCR models, now in the console

You can now choose four more OCR models in the console and through the API: GLM-OCR, LightOnOCR-3, dots.mocr and KDL Frontier Parser nano. They are open models that we run on our own GPUs. Each costs 2 credits a page, like the other preview models.

First results on Persian

We read the first page of our Persian benchmark with each model. dots.mocr got 0.8% of characters wrong, LightOnOCR-3 2.1% and KDL nano 3.3%. GLM-OCR is not made for Arabic script and got more than half wrong. This is one clean, typed page, not a full test, so treat it as a first impression. The full benchmark run is next.

The first page after a quiet spell is slow

To keep costs down, these models switch off when nobody is using them. The first page after a quiet spell waits while the model starts: about 3 to 6 minutes, depending on the model. After that, a page takes 1 to 3 seconds.

So that nobody waits by surprise:

• these models are only used when you choose them; Runemic never falls back to them on its own, and the free tool never uses them
• in the playground, a model starts as soon as you pick it, so it is often ready by the time you have chosen a page
• through the API, the request simply waits until the model has started; it does not fail

How to use them

In the console, pick a model in the playground, or tick up to three to compare them on the same page. In the API, pass the model id, for example preview-dots-mocr:

curl https://api.runemic.com/v1/ocr \
  -H "Authorization: Bearer $RUNEMIC_API_KEY" \
  -F "file=@page.png" \
  -F "model=preview-dots-mocr"

GET /v1/models lists every model with its credits per page. See the API docs for the rest.