A simulator for connectionist networks which uses gradient methods of nonlinear optimization for network learning is described. The simulator (GRADSIM) was designed for temporal flow model connectionist networks. The complete gradient is computed for networks of general connectivity, including recurrent links. The simulator is written in C, uses simple network and data descriptors for flexibility, and is easily modified for new applications. A version of the simulator which precompiles the network objective function and gradient computations for greatly increased processing speed is also described. Benchmark results for the simulator running on the DEC VAX 8650, SUN 3/260 and CYBER 205 are presented.