Chinese researchers say they have achieved a global first in using a real quantum computer to fine-tune an artificial intelligence (AI) model with 1 billion parameters, showing the potential of quantum computing to help better train large language models.
Using Origin Wukong, China’s third-generation superconducting quantum computer with 72 qubits, a team in Hefei has achieved an 8.4% improvement in training performance while reducing the number of parameters by 76%, state-owned Science and Technology Daily reported on Monday.
“This is the first time a real quantum computer has been used to fine-tune a large language model in a practical setting. It shows that current quantum hardware can begin to support real-world AI training tasks,” said Chen Zhaoyun, a researcher at the Institute of Artificial Intelligence under the Hefei Comprehensive National Science Centre.
The improved AI model also reportedly delivered better results on specific tasks. When trained on mental health conversation data, it made 15% fewer mistakes and in a maths problem-solving test, its accuracy rose from 68 to 82%, according to Science and Technology Daily.
Fine-tuning is a key step in customising general AI models such as DeepSeek or Qwen for specialised tasks, such as analysing medical data. Traditionally, this process relies on powerful servers and faces multiple challenges, including limited ability to scale and high energy consumption.
Quantum computing, by contrast, brings unique advantages. By leveraging quantum principles such as superposition – one particle holding multiple possible states at once – and entanglement, which means particles remain linked and instantly affect each other, quantum computers can explore vast combinations of parameters simultaneously, making AI training much faster and more efficient.
To enable this, researchers from Origin Quantum – a Hefei-based start-up that developed the Origin Wukong computer – worked with collaborators to create a new method called quantum-weighted tensor hybrid parameter fine-tuning.
In this approach, the AI model’s settings – called weights – are handled by a network that blends quantum and classical techniques. The quantum part identifies complex patterns in the data, while the classical part compresses the model to use less computing power.
On the Wukong chip, a single batch of AI training data can trigger hundreds of parallel quantum tasks, demonstrating the chip’s ability to handle intensive workloads, according to Dou Menghan, vice-president of Origin Quantum.
Launched in January 2024, Origin Wukong is now among the most advanced programmable and commercially available superconducting quantum computers in the world. More than 80% of its hardware and software components are domestically produced, with in-house backup systems also in place, the company said.
By February, the platform had attracted more than 20 million visits from users in 139 countries and regions, completing hundreds of thousands of tasks in industries such as biomedicine, fluid dynamics and finance.
International users from the United States, Russia, Japan and Canada were among the most active – with US users consistently topping overseas visits.
“US quantum computers are not open to China,” said Guo Guoping, a leading quantum physicist at the University of Science and Technology of China and co-founder of Origin Quantum, in an earlier interview with the South China Morning Post.
“But, adhering to the notion of scientific exploration without borders, we are willing to open our services to users around the world, including the US, to jointly promote the concept of quantum computing for the benefit of mankind,” he said.
The company revealed that a more advanced, fourth-generation superconducting quantum computer, Origin Wukong 2, had entered its final development phase. – South China Morning Post
Source: The Star