The Mississippi Alluvial Plain (MAP) is one of the most important agricultural regions in the United States and underlies about 32,000 square miles of Missouri, Kentucky, Tennessee, Mississippi, Louisiana, and Arkansas. The MAP region supports a multibillion-dollar agricultural industry. The MAP is part of the Mississippi Embayment with several water-bearing units that make up the Mississippi Embayment Regional Aquifer System (MERAS). These water bearing units include the Mississippi River Valley Alluvial aquifer, Claiborne aquifers and Wilcox aquifers. In northeastern Arkansas, the Cache area has been designated as a critical groundwater areas because of decades of groundwater declines that resulted from past and current water use. The objective of the report assocated with this data release is to document and describe the construction, calibration, and results of the inset groundwater-flow model developed for the Cache River Critical Groundwater Area using the latest MODFLOW-6 code. The Cache model derived boundary conditions from the parent MERAS 3 regional model to provide higher resolution simulations in the Cache focus areas. The Cache model was spatially discretized into 500-meter x 500-meter orthogonal cells on a grid. The Cache model had 18 vertical layers, 161 rows, and 142 columns and simulated the Quaternary-age alluvial aquifer with 5-m constant thickness layers and increasing thickness layers for the Tertiary-age units below the alluvial aquifer. The Cache model included 148 stress periods with a simulation period from January 1, 1900 through December 31, 2018 where stress periods: April 1, 2007 through December 31, 2018 where monthly stress periods. Areal recharge was simulated by a soil-water-balance model of the MERAS and passed to the groundwater models. The model simulated agricultural and municipal pumping. The model simulated groundwater-surface water interactions and total streamflow by adding runoff from the soil-water-balance model. The model featured high-dimensional parameterization schemes for calibration using the PEST++ Iterative Ensemble Smoother. Mean absolute residuals for the calibrated priority well observations was 1.58 meters. Mean horizontal hydraulic conductivity for the alluvial aquifer was about 36 meters per day for the Cache model. Calibrated annual areal recharge was 2.2 inches for the Cache model. Primary groundwater outflows represented in the model were from agricultural wells. New avenues for highly parameterized environmental model calibration have become available with computationally efficient ensemble methods. Such methods, however, can be affected by spurious parameter correlation whereby some combinations of parameters may be erroneously adjusted to improve the fit to the observation data when they are spatially and temporally unrelated. Furthermore, ensemble methods can suffer “ensemble collapse” with too many failed runs, limiting the utility of a calibration to find an optimal range of parameter values and uncertainty. A technique called localization can help address both spurious correlation between parameters and ensemble collapse. However, localization is computationally intensive to implement, especially for highly parameterized models. Localization computation time can be substantially longer than hundreds of parallelized ensemble forward runs. Therefore, tradeoffs need to be evaluated that affect the efficiency and perhaps outcome of a modeling analysis, such as between the benefits of limiting spurious parameter correlations, reasonable solution times, efficient use of computing resources, and allowing enough degrees of freedom to adequately explore parameter space. In this study, we compare the results of a MODFLOW6 groundwater-flow model developed for a region with extensive groundwater irrigation in northeastern Arkansas, history matched using 11,746 adjustable parameters and 6,818 observations with and without the auto-adaptive localization feature of the PEST++ iterative ensemble smoother tool PESTPP-IES. Two versions of the same model were run with an initial prior ensemble of 500 realizations and three iterations, whereby 194 runs completed with localization and 192 runs completed without localization. Total calibration runtimes were 15 hours with localization and one hour without localization. Preliminary results indicate that the number of parameters have an impact on the number of completed realizations while the computational burden of localization substantially increased total solution times by an order of magnitude. The average objective function was about 5 percent larger for the runs with localization. Some benefits of localization included better preservation of spatial parameter zones, less parameter deviation from prior values, and less overfitting. Such information is important for developing efficient workflows for ensemble methods, especially when running on pay-as-needed cloud resources.