Emerging applications like Heat-Assisted Detection and Ranging and thermal LIDAR necessitate a shift toward high-resolution, information-rich infrared perception. However, conventional infrared detectors face challenges in area, latency, and power consumption. To address these limitations, we demonstrated a Spintronic Poisson Bolometer that operates in a Poisson-counting regime, where infrared detection is achieved through thermally activated transitions between discrete magnetization states. Building on this platform, we present the concept of a neural bolometer that integrates spintronic poisson bolometers with artificial neural network architectures. Leveraging their tunable stochastic response, nanoscale footprint, and CMOS compatibility, these devices enable in-pixel computation. Our approach enables long-wave infrared (LWIR) detectors to function as both imaging sensors and stochastic computing units. The system achieves LWIR MNIST classification with 92.5% accuracy, 2.82μs inference time, and 1.9μW power consumption. The proposed approach can usher-in a new generation of thermal imagers with in-pixel computing for next-generation of thermal vision.