The productivity of smallholder dairy farms is very low in developing countries. Important genetic gains could be realized using genomic selection but genetic evaluations need to be tailored for lack of pedigree information and very small farm sizes. To accommodate this situation, we propose a flexible Bayesian model for the genetic evaluation of milk yield which allows us to simultaneously account for non-genetic random effects for farms and varying SNP variance (BayesR model). First, we use simulations based on real genotype data from Indian crossbred dairy cattle to demonstrate that the proposed model can separate the true genetic and non-genetic parameters even for small farm sizes (2 cows on average) although with high standard errors in scenarios with low heritability. The accuracy of genomic genetic evaluation increases until farm size is approximately 5. We then apply the model to real data from 4,655 crossbred cows with 106,109 monthly test day milk records and 689,750 autosomal SNPs. We estimate a heritability of 0.16 (0.04) for milk yield and using cross-validation, a genomic estimated breeding value (GEBV) accuracy of 0.45 and bias (regression of phenotype on GEBV) of 1.04 (0.26). Estimated genetic parameters are very similar using BayesR, BayesC and genomic BLUP approaches. Candidate genes nearby the top variants, IMMP2L and ARHGEF2, have been previously associated with milk protein composition, mastitis resistance and milk cholesterol content. The estimated heritability and GEBV accuracy for milk yield are much lower than those from intensive or pasturebased systems in many countries. Further increases in the number of phenotyped and genotyped animals in farms with at least 2 cows (preferably 3-5 to allow for dropout of cows) are needed to improve the estimation of genetic effects in these smallholder dairy farms.