Coalbed methane (CBM) has emerged as a clean energy resource in the global energy mix, especially in countries such as Australia, China, India and the USA. The economical and successful development of CBM requires a thorough evaluation and optimization of well placement prior to field-scale exploitation. This paper presents a two-stage, step-wise optimization framework to obtain the optimal placement of wells for large-scale development of CBM reservoirs. In the first stage, an optimal uniform well pattern is obtained by optimizing well pattern description parameters with the particle swarm optimization (PSO) algorithm. Subsequently, the location and status (active/inactive) of each well are perturbed and optimized within the patterns through the integration of the generalized pattern search (GPS) algorithm and a quality map (QM) representing the production potential. This framework was tested in a synthetic anthracite CBM reservoir in the Qinshui basin (with high gas content and low permeability) and a real field high volatile bituminous reservoir in the Illinois basin (with low gas content and high permeability). The results show that: (i) significant variations in the net present value (NPV) exist with respect to different uniform well patterns (even for cases where the total number of wells are identical), the optima of which can be efficiently determined by the PSO within 100 numerical simulation runs; (ii) the optimization of well perturbations by the GPS results in a more noticeable improvement in NPVs for the synthetic (12.3%) than for the real field model (4.6%); (iii) for the low permeable synthetic model with narrow optimal well spacings (320 m × 200 m), the contribution of the optimization of well perturbation to the NPV increment is heavily dependent on the uniform well placement solution; (iv) for the high permeable real field model with large optimal well spacings (1300 m × 1300 m), the initial uniform well placement has a very minor effect on the subsequent well perturbation solutions in terms of NPV; (v) the proposed framework significantly outperforms the conventional well-by-well concatenation procedure in terms of computational efficiency, robustness and optimal criteria set for production potential.