Fujitsu develops generative AI reconstruction technology for optimized and energy-efficient AI models based on Takane LLM

Fujitsu announced the development of a reconstruction technology for Generative AI (GenAI). This technology is set to enhance the Fujitsu Takane Large Language Model (LLM) by allowing the creation of lightweight, power-efficient Artificial Intelligence (AI) models.

The development features quantization and specialized AI distillation. Fujitsu's proprietary 1-bit quantization achieves a 94% reduction in memory consumption while retaining an 89% accuracy rate compared to unquantized models. This improves inference speed three-fold, allowing large GenAI models to operate efficiently on a single low-end Graphics Processing Unit (GPU).

Fujitsu's specialized AI distillation reduces model size and enhances accuracy, extracting task-specific knowledge for more reliable specialized Artificial Intelligence Security (AIS). This capability facilitates the deployment of advanced AI on edge devices such as smartphones and factory machinery, promoting real-time responsiveness and improved data security while decreasing power consumption.

Fujitsu plans to initiate trial environments for the Takane model with the new technology in the second half of fiscal year 2025, also releasing models of Cohere's quantized research via Hugging Face. Fujitsu aims to advance GenAI capabilities to address complex societal challenges.