>>105654899
The thing is that the new Mistral Small 3.2 uses a new tokenizer with a bunch of additional language tokens, it's unclear if it was trained to actually make use of them. That quant used the old tokenizer.json file from 3.1 (Since Mistral didn't provide one), as well as the old mmproj file (not known if the vision encoder was updated, but it still seems to have issues with NSFW imagery and poses).