Multiple linear regression models with non random regressors in continuous time are considered. The strong consistency of least squares estimates is established under minimal assumptions on the design when the process of errors is a semimartingale satisfying some regularity condition.