Fujitsu unveils GenAI reconstruction tech, making large models lightweight, faster, and power-efficient with record accuracy.
Fujitsu has introduced a GenAI reconstruction technology that makes large models lightweight, efficient, and more sustainable. Integrated into the Takane LLM, this innovation achieves a 94% cut in memory use while maintaining an 89% accuracy retention rate—far higher than the 20% typical of conventional methods.
The breakthrough relies on two core techniques. First, 1-bit quantization compresses parameters, reducing GPU needs from four high-end units to a single low-end device. Second, specialized AI distillation extracts task-specific knowledge, creating compact models that often surpass the performance of the original. Together, these enable a threefold increase in inference speed and make edge deployment feasible on devices like smartphones and factory equipment.
Early trials show practical benefits. In sales negotiation prediction, the distilled Takane model delivered 11 times faster inference and 43% higher accuracy, while using one-hundredth of the parameters. In image recognition, accuracy for unseen objects improved by 10%, marking threefold progress in just two years. These results prove GenAI can be both faster and smarter when carefully reconstructed.
Fujitsu plans to roll out trial environments for Takane with the new quantization technology later in 2025. The company is also releasing quantized models of Cohere’s open-weight Command A on Hugging Face. By enabling secure, energy-efficient AI at scale, Fujitsu positions Takane as a foundation for agentic AI on the edge, where real-time responsiveness and lower power use are critical.
