We developed a prototype of a mobile, high-resolution, multichannel magnetoencephalography ͑MEG͒ system, called babySQUID, for assessing brain functions in newborns and infants. Unlike electroencephalography, MEG signals are not distorted by the scalp or the fontanels and sutures in the skull. Thus, brain activity can be measured and localized with MEG as if the sensors were above an exposed brain. The babySQUID is housed in a moveable cart small enough to be transported from one room to another. To assess brain functions, one places the baby on the bed of the cart and the head on its headrest with MEG sensors just below. The sensor array consists of 76 first-order axial gradiometers, each with a pickup coil diameter of 6 mm and a baseline of 30 mm, in a high-density array with a spacing of 12-14 mm center-to-center. The pickup coils are 6 ± 1 mm below the outer surface of the headrest. The short gap provides unprecedented sensitivity since the scalp and skull are thin ͑as little as 3 -4 mm altogether͒ in babies. In an electromagnetically unshielded room in a hospital, the field sensitivity at 1 kHz was ϳ17 fT/ ͱ Hz. The noise was reduced from ϳ400 to 200 fT/ ͱ Hz at 1 Hz using a reference cancellation technique and further to ϳ40 fT/ ͱ Hz using a gradient common mode rejection technique. Although the residual environmental magnetic noise interfered with the operation of the babySQUID, the instrument functioned sufficiently well to detect spontaneous brain signals from babies with a signal to noise ratio ͑SNR͒ of as much as 7.6:1. In a magnetically shielded room, the field sensitivity was 17 fT/ ͱ Hz at 20 Hz and 30 fT/ ͱ Hz at 1 Hz without implementation of reference or gradient cancellation. The sensitivity was sufficiently high to detect spontaneous brain activity from a 7 month old baby with a SNR as much as 40:1 and evoked somatosensory responses with a 50 Hz bandwidth after as little as four averages. We expect that both the noise and the sensor gap can be reduced further by approximately half with a gain in SNR of about four. Thus, we conclude from the performance of the prototype that it should be feasible to improve the babySQUID to detect cortical activity in infants in real time with high spatial resolution.