We describe different methods for estimating the bispectrum of Cosmic Microwave Background data. In particular we construct a minimum variance estimator for the flat-sky limit and compare results with previously-studied frequentist methods. Application to the MAXIMA dataset shows consistency with primordial Gaussianity. Weak quadratic non-Gaussianity is characterised by a tunable parameter $f_{NL}$, corresponding to non-Gaussianity at a level $∼ 10^{-5}f_{NL}$ (ratio of non-Gaussian to Gaussian terms), and we find limits of $|f_{NL}|<950$ for the minimum-variance estimator and $|f_{NL}|<1650$ for the usual frequentist estimator. These are the tightest limits on primordial non-Gaussianity which include the full effects of the radiation transfer function. Comment: 24 pages, 13 figures