Which model can you recommend for local processing with Ollama

[mikedolx]

Hi,

I'd like to ask you, which model you could recommend for processing documents on a locally hosted ollama instance. I have assigned a 3060 with 12GB RAM to my Ollama instance.

I migrated all of my documents from my old DMS into paperless and let Ollama with Phi4:latest do some recognition over night. The results were OK, but I think it could be better.

[kmbanana]

Also using a 12GB 3060 and using Phi4 with "OK" results. I had high hopes for gemma3:12b, but got a lot of responses back in German when the documents were clearly English and I'm not sure why. If I use the same model and copy the default paperless-AI prompt exactly and put them into open-webui I get English results from the same documents paperless-AI tagged with German names/tags. Ollama had some issues with Gemma so maybe I'll come back and try again in another month.

[troubleshootme]

From my experience, Phi4:14b has been the best so far.