Sustainable Intensification (SI) interventions are urgently required, particularly those suited to resource poor farms in Africa. Legume crops have been promoted as a key ingredient for SI of rainfed grain production on smallholder farms, with variable results. There is a need to explore the extent to which legume diversification of maize production impacts stability, nutrition, and income. This is particularly so for marginal environments. On-farm experimentation was conducted in Central Malawi over two to seven years on 29 farm sites (120 year-site combinations). The farms were located within four areas that included lakeshore, dissected hills (two locations) and upland plain. Maize diversification included a long-duration legume pigeonpea and a medium duration legume groundnut, grown in rotation or intercropped with maize, and as a doubled-up legume rotation (DLR). To quantify the performance of systems under low, medium, and high yield environments, we used long-term average maize yield to categorize each farm site. All legume diversified systems supported stable grain production in the low yield environment, as shown by 37-41% coefficient of variation for yield, in comparison to 62% for sole maize. The groundnut systems consistently produced the highest grain yield, protein, stable yields, and economic returns, and this performance held up in marginal, low yield environments. In this multi-site, multi-year, on-farm replicated study, the performance of groundnut systems (GnRot and DLR) stood out for high protein (0.529 T ha-1 2 yr-1 and 0.615 T ha -1 2 yr-1 , respectively over two years) versus unfertilized maize (0.169 T ha-1 2 yr-1). These two groundnut-based systems were produced with half-fertilizer rates compared to sole maize and were economically high performers. However, there was a barrier to adoption of GnRot and DLR in that improved groundnut seed was expensive (USD 157 ha-1 2 yr-1), this initial investment being beyond the means of many farmers, despite the cost largely offset by the generation of high income (USD 1636 to 1993 ha-1 2 yr-1). Long-term sustainability was assessed by monitoring soil organic carbon (SOC), which was found to be markedly influenced by soil texture (sites with SOC > 1.5% had sand content <50%). Legume diversification effects on SOC were not discerned, possibly due to high sand content on the oldest trial sites. This study highlights the value of longitudinal data and including a wide range of soil texture sites in on-farm experimentation to identify overall legume diversification effects within maize systems.