The Roman Space Telescope is expected to increase samples of cosmologically useful Type Ia supernovae (SNe Ia) by an order of magnitude in the next decade, leading to unprecedented constraints on the properties of dark energy. However, current measurements are limited by uncertainties in the modeling of SNe Ia, and many SN Ia models are restricted to ultraviolet and visual wavelengths. To address this we’ve created a new training methodology and model framework, which we call SALT3. Our SALT3.K21 model has been trained on 2.5 x as many SNe, and the potential impact of calibration uncertainties has been greatly reduced as compared to the widely-used SALT2.4 model. Our model has an extended wavelength range 2000-11000 angstroms (1800 angstroms redder) and reduced light-curve uncertainties, enabling accurate use of low-z I and iz photometric bands. We’ve shown that SALT3.K21 reduces statistical uncertainty in the low-z Foundation and CfA3 SN samples by 15% and 10%, respectively, by taking advantage of this previously discarded data. To check for potential systematic uncertainties we compare distances of low and high redshift SNe in the training compilation, finding an insignificant 2(14) mmag shift between SALT2.4 and SALT3.K21. Based on simulations, SALT3.K21 is suited to make use of 95% of the SN Ia epochs projected to be observed by the Roman Space Telescope, 20% more than SALT2. Ongoing development is focused on creating a model that extends from the optical into the near infrared in anticipation of increased low-z NIR SN Ia samples to be produced by Roman.