The relationship between length (L) and weight (W) was estimated for 80 species belonging to 50 families of marine fishes from the shelf and upper slope of southern Brazil (lat. 28°S - 34°S). Sample sizes (n) for different species ranged from 11 to 14 741 specimens collected from commercial landings and research surveys. The fit of the equations (W=aLb) with a and b parameters estimated from regular and functional regression (of log-transformed weight and length data) as well as from a non-linear iterative process using the quasi-Newton algorithm were compared. The non-linear method gave the most accurate estimates in terms of residual sum of squares. Differences were less than 2.3% for n>500 compared with predictive regressions and 1.5% compared with functional regressions. No difference was observed between both predictive and functional regressions. Determination coefficients (r2) increased with sample size, and the highest r2 were obtained for 50