The equation for radiative transfer is $$\omega \cdot \nabla I = \kappa(B - I)$$ where $I$ is the intensity of radiation, $\omega$ is the ray direction, $\kappa$ the absorption coefficient, $B$ the Planck function. Here, $\kappa=\kappa(\nu)$ (i.e., it depends on the radiation frequency $\nu$).

Often, though, the gray model, where $\kappa$ doesn't depend on $\nu$, is used. How can this be justified? What assumptions are used to obtain the gray model from the non-gray model?

-
I've modified your question to be more clear in hopes to reopen it. I believe that I kept the intent of your question in tact, but please check this so that we are answering the question you want answered. –  Kyle Kanos Jul 7 at 2:38
@KyleKanos Thank you. –  jokersobak Jul 7 at 8:25

The light ray from the source (i.e., the star's atmosphere) travels at some angle, $\theta$, from normal, $z=0$. Since the light is coming from an angle, we need to account for that by modifying the radiative transfer equation to have a vertical optical depth, defined by $$\tau_{\lambda,v}(z)=\int_z^0\kappa_\lambda\rho\,dz$$ which gives us $$\omega\frac{dI_\lambda}{d\tau_{\lambda,v}}=I_\lambda-B_\lambda$$ with $\omega=\cos\theta$. Since the path length of the ray does not offer a unique geometric depth of the atmosphere, we cannot use $\nabla I$ and must use the above form for the radiative transfer equation.
In most stellar atmospheres, the primary source of opacity is the photoionization of H$^-$ ions. Fortunately, the opacity from photoionzation of hydrogen ions does not vary rapidly with wavelength, so we can approximate the opacity as $$\kappa_\lambda\approx \bar\kappa$$ where $\bar\kappa$ is some average opacity, usually the Rosseland mean opacity.