Fujitsu aims to contribute to a sustainable AI society by making AI more accessible, efficient, and secure, aligning with its broader mission to build trust through innovation.

This breakthrough allows large generative AI models to operate efficiently on low-end GPUs and edge devices like smartphones and factory machinery, enabling real-time responsiveness and enhancing data security.

Demonstrations have shown an 11-fold increase in inference speed and a 43% improvement in accuracy for sales prediction models, with smaller student models outperforming larger teacher models.

Starting in the second half of fiscal year 2025, Fujitsu plans to offer global trial environments for Takane using this technology and has begun releasing quantized models via Hugging Face, including models from Cohere.

Future plans include further reducing model sizes up to 1/1000 of their current size while maintaining accuracy, and developing specialized AI for sectors like finance, healthcare, manufacturing, and retail, ultimately evolving into autonomous AI systems for complex problem-solving.

Fujitsu also announced a new generative AI reconstruction technology that creates lightweight, energy-efficient models, further enhancing their Kozuchi AI service and Takane Large Language Model.

Fujitsu has developed a groundbreaking AI technology that combines 1-bit quantization and specialized AI distillation, drastically reducing model size and power consumption while maintaining high accuracy and tripling inference speed.

The technology mimics human brain self-reconstruction by reorganizing neural circuits, extracting only necessary knowledge to create lightweight, reliable, and highly efficient AI models.

In image recognition, the technology achieved a 10% accuracy improvement over existing distillation techniques, marking a significant breakthrough over the past two years.

This innovation supports Fujitsu’s commitment to the UN Sustainable Development Goals by promoting sustainability, trust, and global impact through advanced AI.

This technology employs advanced quantization algorithms and brain-inspired optimization techniques such as pruning and neural architecture search to prevent error growth and enhance model efficiency.

By compressing neural network connections, the technology enables large models to run efficiently on low-end GPUs, making advanced AI more accessible and sustainable.