Miniaturized near infrared spectrometers are now available, at more affordable prices than conventional spectrometers, but their performances have been poorly studied to date. This paper aimed at comparing the performances of the JDSU MicroNIR 2200 spectrophotometer (weight < 0.1 kg) with those of a conventional bench-top instrument for predicting carbon and nitrogen contents in laboratory conditions, on a range of representative Malagasy soils. Though its noticeably narrower and less resolved spectra (1151-2186 nm at 8.15 nm step vs. 1100-2498 nm at 2 nm step), the microspectrometer yielded predictions in independent validation that were almost as accurate as those of the conventional instrument (standard errors of prediction were 4.6 vs. 3.4 gC kg(-1), but 3.9 vs. 3.4 gC kg(-1) after bias correction, and 0.36 vs. 0.35 gN kg(-1), respectively). Due to noisy features, the MicroNIR spectra needed mathematical pretreatment (e.g. standard normal variate SNV), and bias correction for C, for providing accurate predictions, while the raw absorbance spectra from the conventional instrument did not. Furthermore, building multivariate models with MicroNIR spectra required less latent variables than with their conventional counterparts, and these models were less prone to performance degradation when applied to independent validation samples. Fitting the spectra of the conventional instrument to those of the MicroNIR (1150-2182 nm at 2 or 8 nm step) showed that (moderately) less accurate MicroNIR predictions could be firstly attributed to narrower spectral range rather than to poorer resolution. Considering their performances, such microspectrometers could thus represent a cost-effective alternative to conventional spectrometers. They have now to be tested in field conditions.