Best vision model

[mvnixon]

Hi!
What is the consensus on the 'best' Ollama vision model? Right now im using minicpm-v but I'm unhappy with the performance. Llava didn't work when I tried, anything else worth trying?

[Psypher37]

Hi there,

I'm currently having a lot of success with the qwen2.5vl:7b model running on a Nvidia 4070Super. It's much better than the minicpm-v.

I also tried granite3.2Vision but found it unsatisfactory. Mistral-small3.1 was also very good, but I could only run it via CPU which made it unreasonably slow to process multiple documents.

I'm also using mistral-nemo for the suggestions - it follows instructions well as long as the prompts are prescriptive. The OCR part is pretty important though - garbage in garbage out if the OCR is not great.

[ManuVice]

did you try qwen3 yet?
whats your average time for a document with qwen2.5 ?

[Psypher37]

No I haven't yet. I switched to a quantized version of Mistral-small3.1 which let me run it on my GPU instead. It's not as fast as qwen but I found it to be more accurate.

If you want to suggest a public PDF, I could run it a few times on different models & give you the processing times, but honestly it's what works for you. I'm not too fussed about longer processing times given that I'm only adding a few documents at a time.