I work on computing VaR for equity derivative portfolios in a bank. My neighbour is working on a prototype engine to tackle large deviations. Unfortunately, this is still based on Gaussian Copulas and "sum of gaussian approximations", because gaussians enable nice linear parametric VAR computations which is:

1. about the only one we can compute daily with existing computing power (I guess this is the same everywhere: it's just one giant matrix multiplication, and this is very efficient)
   
2. the only one that you can explain nicely by projecting the VAR on the risk factors (otherwise, you cannot have a "dashboard" to steer the activity of your traders in order to contain your risk within set bounds).

And he still has a lot of trouble with his models. Like mig says, there are few large deviations to calibrate the stuff, and Gauss does indeed work well day to day (as proven by backtesting over periods without krachs). It's only when things get real bad that you see your quantile was wrong (about to know pretty soon I guess).

Mandelbrot claimed that scaling laws where better than Gauss, can be calibrated, and incur manageable computing costs, but I don't think anyone has looked into it on the industry-scale. I'll dig a link

Pierre