Nord v4.2: Brain-Inspired Spiking Neural Network Language Model with Emergent Zonal Specialization at 618M Scale

Large language models (LLMs) based on the Transformer architecture have demonstratedremarkable capabilities in natural language processing. However, these models activate 100% ofparameters for every input token, leading to high computational and energy costs. Spiking NeuralNetworks (SNNs), inspired by biological neural computation, offer a fundamentally differentapproach: neurons communicate through discrete binary spikes, and most neurons remain silentmost of the time.Several SNN language models have been proposed. SpikeGPT (Zhu et al., 2023) demonstratedlanguage generation at 216M parameters using an RWKV-based architecture. BrainTransformers(LumenScope, 2024) achieved competitive benchmark scores at 3B parameters using anANN-to-SNN training pipeline. SpikeLLM converts pretrained LLaMA weights to spiking form.However, none of these architectures exhibit emergent functional specialization of architecturalzones during training.Nord v4.2 introduces a zonal SNN architecture where Sensory, Association, Memory, and Executivezones develop functionally distinct firing rate patterns from uniform initialization through standardgradient-based training with spike homeostasis regulation. This emergent self-organization mirrorsbiological cortical organization, where prefrontal cortex exhibits higher baseline firing rates thanprimary sensory cortex (Mountcastle, 1997).