We present an improved algorithm for selecting variable precision coefficients for FIR filters that produces a reduced space implementation with no degradation in frequency response. The algorithm is based on the fact that the frequency response of a filter has different sensitivities to different coefficients depending on the response itself (i.e. the coefficient value). The method provided here makes it possible to predict the variable precision of the quantized coefficients that are required to meet the specification. This approach, along with other technologies such as CSD and the scaling method used in our realizations, has opened exciting implementation possibilities for FPGAs, ASICs, and custom VLSI. The example study in this paper shows that using variable precision to exploit redundancy across the coefficients results in significant reductions in complexity and area over the uniform wordlength method.