Abstract: As a typical initial and boundary-value problem, data assimilation technique plays a very important role in numerical weather prediction （NWP）. In the past decades, the 3D and 4D-Var data assimilation techniques were improved quickly not only in the theoretical researches but also in the NWPs of operational centers. However, there are still some problems which limit the wide use of variational data assimilation techniques. One of them is that the general 3-dimensional variational data assimilation （3DVar） technique is short of complicated constraints such as the dynamics and physics in a numerical model, as are used in the 4-dimensional variational data assimilation （4DVar） technique. The other one is that using a numerical model and its adjoint with the 4DVar technique requires a large amount of computer resources, and thus limits its practical applicability. A new 3DVar method (mode-constrained 3DVar, MC-3DVar) is then proposed by incorporating a numerical model constraint. This method minimizes the distance between the observations and model variables and also their time tendency, so that the optimized initial conditions not only fit the observations but also satisfy the constraints of full dynamics and physics of the numerical model. In this study, an assimilation cycle is employed to improve the initial conditions using various data at different time. A case study of Typhoon Saomai （2006） from 20：00 BST 8 to 05：00 BST 9 August is carried out. The assimilated data include the cloud drift wind, QuikSCAT sea level wind, and Bogus sea level pressure. The track forecast of Saomai is improved dramatically after assimilating these data. The simulated structure of the typhoon is also improved using the new initial conditions after the assimilation cycle. These can be attributed to that using MC-3DVar method, after the Bogus sea level pressure field is assimilated, the height, wind and other variables are also adjusted by the constraints of the dynamics and physics of the numerical model. In addition to the Bogus vortex data, the other assimilated observations are also important to improve the initial conditions.