Use as many data as you have. Also, nonlinear parameter estimation techniques are very sensitive to the initial estimates (that you give to the routine to start with), and an inaccurate set can cause the routine to end up in a local minimum rather than a minimum that is much closer to the correct parameters. Choosing the correct values can be challenging.
If you repeatedly have problems guessing the correct initial parameter values, use one of the Global Optimization Toolbox functions (such as the genetic algorithm ga function) to search out the best parameter set. Those take time, however they are usually succesful. (For ga, begin with a large initial population, so it has a better probability of discovering the best parameter set.)