Fujitsu has developed a groundbreaking reconstruction technology for generative AI, designed to enhance its Fujitsu Kozuchi AI service. This innovation will strengthen the Fujitsu Takane large language model (LLM) by enabling the creation of lightweight, energy-efficient AI models capable of running on minimal hardware.
This breakthrough is driven by two major innovations: quantization and specialized AI distillation. Using Fujitsu’s own 1-bit quantization technology on Takane, the team was able to reduce memory usage by an impressive 94% all while maintaining nearly 90% of the accuracy of the original model. This improvement translates into a three-fold increase in inference speed, significantly outperforming conventional quantization approaches. Thanks to this, large generative AI models that used to need several high-end GPUs can now run smoothly on just a single, modest GPU.
AI Authority Trend: Fujitsu, 1Finity, and Arrcus Partner to Strengthen AI-Era Network Infrastructure
On top of that, Fujitsu has rolled out a groundbreaking, first-of-its-kind AI distillation technique. This approach not only drastically reduces model size but also boosts accuracy beyond the original model’s capabilities. Drawing inspiration from the human brain, this method extracts and condenses task-specific knowledge, creating highly efficient and reliable specialized AI models.
“This revolutionary lightweighting capability will democratize access to advanced AI technologies,” Fujitsu stated. “It enables sophisticated agentic AI to operate seamlessly on edge devices such as smartphones and industrial machinery, improving real-time responsiveness while safeguarding data and significantly reducing energy consumption.”
AI Authority Trend: Fujitsu’s 1FINITY Is Here: A New Era of Smarter, Faster Networks Begins
By making AI models more compact and efficient, Fujitsu aims to support sustainable AI practices and facilitate wider deployment across various sectors. This includes applications in manufacturing, healthcare, and mobile devices, where low-latency processing and secure data handling are critical.
Fujitsu plans to launch trial environments for Takane incorporating this technology in the second half of fiscal year 2025. The company will also gradually release models of Cohere’s open-weight Command A, quantized through Hugging Face. “We remain committed to advancing generative AI capabilities to address complex societal challenges and unlock new opportunities for AI utilization,” Fujitsu added.
With these innovations, Fujitsu is setting a new benchmark for generative AI efficiency, enabling powerful, sustainable, and accessible AI solutions across devices and industries worldwide.
AI Authority Trend: Fujitsu’s Uvance Wayfinders Helps Businesses Transform with Data and AI
To share your insights, please write to us at sudipto@intentamplify.com
