Great work. One would expect loop regions to be more flexible anyway so you are certainly on the right track. I think ERRAT and Verify3D also state a min % to expect with a good model (I think it is 90%). The main problem is your alignment percentage. Without increasing your alignment % you are unlikely to achieve significantly better results. I think that is where your focus should be. Try yamules suggestions above or post a question recommending aligners for homology modelling.
Also, are any of the loops in structurally/functionally significant locations? Check Swiss-prot
Alternatively, while Swiss-model would be my choice of online modelling tools, I would recommend using Modeller instead locally if you can install it.
RAMPAGE has also been suggested but this is just another tool for producing (nicer) ramachandran plots
Also check out this post for further refinement
Your result isn't bad considering you had only 33% alignment (the most important determinant of the model).
Next steps: Are you sure the best you can find is 33%? Have you considered relaxing the BLAST parameters? Homology modeling is a model-->refine>>model>>refine process. Is your template from NMR or X-ray? - Is the template high quality? (post the link to the PDB entry here please) What do the ERRAT and Verify3D results look like? - what are the problem regions? Have you visualised the model mapped on to the template? Where do they differ?
How about using ROBETTA http://robetta.bakerlab.org/ or I-TASSER http://zhanglab.ccmb.med.umich.edu/I-TASSER/ ? Or using HMM-HMM alignment for sequence alignment. It is more accurate than the alignment by BLAST.