Mitsubishi Electric has developed a language model tailored for manufacturing processes operating on edge devices.

The Maisart-branded AI technology has been pre-trained with data from Mitsubishi Electric’s internal operations. This enables it to support a wide range of applications in specific manufacturing domains. The language model employs a proprietary data-augmentation technique to generate responses optimised for user-specific applications.

Mitsubishi argues that the adoption of large language models (LLMs) for generative AI are raising concerns over their significant computational and energy costs.

The company also suggests that there is an increasing demand for generative AI solutions that can operate in on-premises environments due to data privacy and confidential information management requirements.

As a response, Mitsubishi Electric has developed a domain-specific language model by training a publicly available Japanese base model with the company’s proprietary data from its own business domains, including factory automation (FA).

Using training data generated through the company’s original augmentation techniques allowed task-specific fine-tuning. The resulting model is compact enough to run on limited hardware resources, making it suitable for environments with constrained computing capabilities such as edge devices, as well as for on-premises operations such as call centres that handle sensitive customer information.

The language model was developed as part of the AWS Japan Generative AI Accelerator Programme, a Japan-specific support programme developed by AWS Japan in 2023, building on a large-scale language model (LLM) focused development support programme, and launched in July 2024.