Spiking Neural Networks for Edge Intelligence

We explore how event-driven SNNs can replace transformer attention for low-power edge inference, achieving competitive performance at a fraction of the energy cost.

Introduction

Spiking Neural Networks (SNNs) represent a fundamentally different approach to neural computation. Unlike conventional artificial neural networks that pass continuous values through layers synchronously, SNNs communicate through discrete spike events — just like biological neurons.

Why SNNs for the Edge?

Modern large language models require gigawatts of compute just for inference. At Korollr, we believe this is fundamentally unsustainable. SNNs offer a path forward: sparsity by design, where neurons only activate when they receive enough input to spike. Event-driven processing, where no computation happens when there's no activity. And temporal encoding, where information is carried in spike timing, not just values.

Lambert's Architecture

Our Lambert series models incorporate SNN-inspired components in the attention mechanism. We're developing threshold-based sparse attention that activates only for relevant token pairs, replacing full attention matrix computation with event-driven alternatives. This is an active area of research — our current models already support SNN conversion, and we're iterating toward fully native spiking inference.

Results

In our internal benchmarks, Lambert Lemma 0.2 runs at 20–40 tokens/second on consumer CPUs. For everyday conversational tasks, it performs better than its size suggests — but we'd rather you test it yourself than take our word for it.

Next Steps

We're actively exploring full SNN replacements for the feed-forward blocks. The goal: a model that consumes less than 1W during inference on custom neuromorphic silicon.