This research redesigns large language models (7-70B parameters) using bio-inspired spiking neural networks, achieving substantial efficiency gains while maintaining performance.

• Introduces saliency-based spiking mechanism that mimics human brain efficiency
• Achieves up to 62% reduction in computational costs during inference
• Demonstrates scalability across multiple LLM architectures (Llama, Mistral, Qwen)
• Maintains 96-99% of original model performance while reducing energy requirements

This engineering breakthrough addresses one of the most critical challenges in AI today: reducing the massive computational resources needed for LLM inference, potentially enabling more sustainable and accessible AI deployment.

SpikeLLM: Scaling up Spiking Neural Network to Large Language Models via Saliency-based Spiking