Recently, a research team led by Guoqi Li and Bo Xu from the Institute of Automation, Chinese Academy of Sciences, published a paper titled "Neuromorphic Spike-based Large Language Model" in National Science Review. The study introduces a novel neuromorphic spike-based large language model (NSLLM), which draws inspiration from principles in neuroscience to significantly enhance the energy efficiency and interpretability of large language models (LLMs). This breakthrough not only opens a new pathway for developing efficient artificial intelligence but also offers valuable insights for the design of next-generation neuromorphic chips. The research is a collaborative effort involving multiple institutions across China and abroad, including the Tianqiao and Chrissy Chen Institute for Brain Science, Beijing Academy of Artificial Intelligence, Beijing Zhongguancun College, University of California, Tsinghua University, Peking University, Luxi Technology, University of Sydney, Hong Kong Polytechnic University, AMD, University of Chinese Academy of Sciences, and Ningbo University. The findings have been published in National Science Review. LLMs have become a cornerstone in the pursuit of artificial general intelligence (AGI). However, their widespread deployment brings substantial computational and memory costs, limiting their potential as foundational infrastructure for society. Moreover, existing LLMs often lack transparency—making their decision-making and optimization processes opaque—hindering their reliability and fairness in high-stakes domains such as healthcare and finance. In contrast, the human brain performs complex cognitive tasks with less than 20 watts of power and exhibits remarkable clarity in information processing. This stark contrast highlights two critical challenges: improving the energy efficiency of LLMs and enhancing their interpretability. To address these challenges, the research team developed a unified interdisciplinary framework that bridges neuroscience and LLMs. By implementing integer spike counting, binary spike conversion, and spike-based linear attention mechanisms, the team transformed conventional LLMs into NSLLMs. This innovation enables direct application of neuroscience tools to analyze the model’s internal information processing dynamics. A key advancement is the introduction of an "integer training, binary inference" paradigm. This approach converts standard LLM outputs into spike sequences, enabling the use of neuroscientific methods to examine decision-making processes. As a result, the model’s logic becomes more interpretable. To validate energy efficiency, the team designed a MatMul-Free hardware architecture on an FPGA platform for a 1-billion-parameter NSLLM. Through layer-wise quantization and sensitivity analysis, they optimized a mixed time-step spiking model, achieving competitive performance with low-bit precision. By incorporating a quantization-aided sparsity strategy, they adjusted membrane potential distributions to shift spike probabilities toward lower integer values, significantly reducing spike rates and boosting efficiency. On the VCK190 FPGA platform, the team implemented a MatMul-Free core that completely eliminated matrix multiplication operations. The result was a dynamic power consumption of just 13.849W, with a throughput of 161.8 tokens per second. Compared to an A800 GPU, this approach achieved 19.8× higher energy efficiency, 21.3× better memory efficiency, and 2.2× higher inference throughput. The NSLLM framework also enables enhanced interpretability through neurodynamics analysis. By modeling the LLM’s behavior as a spiking neural network (e.g., spike trains), researchers can study neuron dynamics—such as randomness measured by Kolmogorov-Sinai entropy—and information processing via Shannon entropy and mutual information. Results show that in processing unambiguous text, the model encodes information more effectively. For example, intermediate layers exhibit higher normalized mutual information when handling ambiguous text, while the AS layer displays unique dynamics indicating a role in sparse information processing. The FS layer shows higher Shannon entropy, suggesting stronger information transmission capacity. The positive correlation between mutual information and Shannon entropy further indicates that layers with higher information capacity better preserve critical input features. These findings demonstrate that the NSLLM framework provides biologically plausible explanations for LLM mechanisms while reducing data requirements. Inspired by the brain’s sparse, event-driven computation, the team has created a cross-disciplinary model that replaces traditional LLMs with a neuromorphic alternative. The NSLLM maintains performance comparable to state-of-the-art models in complex tasks such as commonsense reasoning, reading comprehension, world knowledge QA, and mathematical problem solving. This work advances the frontier of efficient AI, offers a new lens for understanding LLM interpretability, and paves the way for future neuromorphic hardware design.