Comment by Iolaum

As long as they are transparent about what quant they serve the model and any other optimization they do that also affects performance of inferred tokens.