SpikingBrain 1.0 mimics how the human brain fires only the neurons it needs, saving power and speeding up the response time, researchers say. — SCMP
A Chinese team has unveiled what it calls the world’s first “brain-like” large language model – an artificial intelligence system designed to use less energy, perform better and operate without Nvidia chips.
Developed by researchers at the Chinese Academy of Sciences’ Institute of Automation in Beijing, SpikingBrain 1.0 mimics how the human brain fires only the neurons it needs.
So instead of activating an entire network like ChatGPT and other mainstream AI tools, it selectively responds to input, saving power and speeding up response time.
Thanks to this design, the model can learn from just a sliver of training data – less than 2% of the amount conventional systems need – while staying fast and efficient even when processing long text.
In some cases, it ran up to 100 times faster than traditional models, according to a non-peer reviewed technical paper posted on arXiv, an open-access research repository.
The system runs entirely on China’s homegrown AI ecosystem, powered by the MetaX chip platform rather than Nvidia’s dominant GPU hardware. That makes the model strategically important as the US tightens export controls on advanced AI chips.
Li Guoqi, a lead researcher at the institute, said the model opened a new path for AI development while delivering a framework optimised for Chinese chips. He said it could be useful to process long sequences of data such as legal documents, medical records or scientific simulations.
Li’s team has open-sourced a smaller version of the model and made a larger one available online for public testing.
“Hello! I’m SpikingBrain 1.0, or ‘Shunxi’, a brain-inspired AI model,” the system says on its demo site. “I combine the way the human brain processes information with a spiking computation method, aiming to deliver powerful, reliable, and energy-efficient AI services entirely built on Chinese technology.”
Today’s most popular AI models, including ChatGPT, require enormous computing power. To train them, companies rely on massive data centres packed with high-end chips that burn through electricity and cooling water.
Even after training, these systems remain resource-hungry. Handling long inputs or generating complex responses can slow them down and strain memory, as they process every word in parallel rather than focusing only on what matters – driving up the cost and environmental impact of running them.
In contrast, the team behind SpikingBrain 1.0 took inspiration from how real neurons work. Rather than processing everything at once, the system reacts selectively, using less power to perform complex tasks – much like the human brain.
Its core technology, known as “spiking computation”, mimics the brain’s habit of firing quick bursts of signals only when triggered. This event-driven approach keeps the system quiet most of the time, helping it stay lean and energy-efficient.
To prove their concept, the team built two versions of the model: one with 7 billion parameters, the other with 76 billion. Both were trained not on Nvidia chips, but on China’s home-grown MetaX platform, built by Shanghai-based MetaX Integrated Circuits Co.
Despite using only a fraction of the data consumed by conventional models – about 150 billion tokens in total – SpikingBrain performed on par with popular open-source alternatives, the researchers reported.
It also excelled at handling long sequences. In one test, the smaller model responded to a 4 million-token prompt more than 100 times faster than a standard system. The team said their set-up ran stably for weeks on hundreds of MetaX chips while consuming far less power.
“These results not only demonstrate the feasibility of efficient large-model training on non-Nvidia platforms, but also outline new directions for the scalable deployment and application of brain-inspired models in future computing systems,” they wrote. – South China Morning Post
