1. We have built a realistic 24-neuron model based on data from the spinal pattern generator for swimming in Xenopus embryos with the use of the SWIM programs. The neurons have dendrite, soma, and axon compartments with voltage-gated Na+ and K+ channels. Dendritic synapses were modeled as modulated ionic conductances with currents that have different reversal levels. One of these conductances was voltage dependent to model N-methyl-D-aspartate ("NMDA") synapses in the presence of Mg2+. 2. In this model, rhythm generation is initiated by a brief excitation, depends on rebound from reciprocal inhibition, and is sustained by long-duration "NMDA-dependent" feedback excitation. 3. Without NMDA voltage dependency, rhythmic activity is stable over a wide range of synaptic conductances. Its frequency decreases with more inhibition and increases with more excitation. The introduction of normally distributed variation in soma size or excitatory synaptic conductance extends the lower stable frequency range. Without such variation the frequency of the 24-neuron model is the same as a 4-neuron model provided that the synaptic conductances for each neuron are the same. 4. The effect of introducing NMDA voltage dependency on rebound after negative current injections or synaptic inhibition was investigated in single depolarized model neurons. With NMDA voltage dependency, hyperpolarizations and rebound spike responses were increased. 5. Network activity with NMDA voltage dependency was similar to that without it, but inhibitory postsynaptic potentials (IPSPs) and spikes were larger, and frequencies were lower and more sensitive to changes in excitatory and inhibitory conductance. 6. We conclude that in the model, mutual reexcitation among excitatory spinal interneurons can sustain rhythm generation by positive feedback and that NMDA voltage dependency can enhance postinhibitory rebound, stabilize swimming activity and extend its lower frequency range, and steepen the dependency of frequency on synaptic drive.