Chinese researchers from the Chinese Academy of Sciences have unveiled SpikingBrain1.0, described as the world’s first “brain-like” large language model (LLM). The model is designed to consume less power and operate independently of Nvidia GPUs, addressing limitations of conventional AI technologies.
Existing models, including ChatGPT and Meta’s Llama, rely on “attention,” a process that compares every word in a sentence to all others to predict the next word. While effective, this approach consumes large amounts of energy and slows processing for long texts, such as books.
Traditional models also depend heavily on Nvidia GPUs, creating hardware bottlenecks for scaling.
Brain-inspired approach
SpikingBrain1.0 uses a localized attention mechanism, focusing on nearby words rather than analyzing entire texts. This mimics the human brain’s ability to concentrate on recent context during conversations. Researchers claim this method allows the model to function 25 to 100 times faster than conventional LLMs while maintaining comparable accuracy.
The model runs on China’s homegrown MetaX chip platform, eliminating reliance on Nvidia GPUs. It selectively responds to input, reducing power consumption and enabling continual pre-training with less than 2% of the data needed by mainstream open-source models. The researchers note that in specific scenarios, SpikingBrain1.0 can achieve over 100 times the speed of traditional AI models.
The development of SpikingBrain1.0 comes amid U.S. technology export restrictions that limit China’s access to advanced chips required for AI and server applications. These restrictions have accelerated domestic AI innovation, with SpikingBrain1.0 representing a step toward a more self-sufficient AI ecosystem.
