ASUS Subsidiary TWSC Release Large Language Model for Taiwanese Culture and Traditional Mandarin Applications

FFM-Llama2 will empower enterprise AI applications, providing support for model fine-tuning and inference

  • Democratizing AI to empower businesses: TWSC offers open-source large language models, AI supercomputing, and AI tools to accelerate AI 2.0 adoption
  • Enhanced traditional Chinese semantics: FFM-Llama2 supports development of Code Llama and embedding models that enhance semantic search
  • Enabling generative AI solutions: FFM-Llama2 is the latest addition to a portfolio of AI solutions that prioritize reliability and security

ASUS today announced that its subsidiary Taiwan Web Service (TWSC), Taiwan’s leading AI technology company has released FFM-Llama2, the world’s first large language model with advanced capabilities in traditional Chinese.

FFM-Llama2 is trained on Llama 2, an open-source large language model created by Meta. FFM-Llama2 harnesses AIHPC supercomputing, parallel computing, and local language data to elevate its proficiency in traditional Chinese. The model is designed to support fine-tuning and customization, empowering enterprises to deploy generative AI solutions.

TWSC has also recently introduced a new Parameter-Efficient Fine-Tuning (PEFT) feature to its enterprise-grade AI Foundry Service (AFS). AFS is a comprehensive set of tools and services that helps businesses develop and deploy custom large language models. The new PEFT feature allows for improved price-to-performance ratios, which will help further drive the adoption of AI 2.0 technologies across business landscapes.

Democratizing AI to empower businesses

TWSC is committed to providing reliable open-source large language models while also continually expanding its AI supercomputing capabilities and perfecting its AI tools platforms. This strategic approach aims to eliminate entry barriers for enterprises and accelerate the widespread application of AI 2.0 technologies.

“Over the past three decades, Taiwan’s ICT industry has thrived within an open framework, establishing the island as a major player in global technology innovation,” says Peter Wu, CEO of ASUS Cloud and TWSC. “We hope to leverage the advances of Taiwan’s ICT industry with our partners to develop AI-powered solutions that enterprises and other organizations can use to enhance business, enriching the lives of people everywhere.”

Enhanced traditional Chinese semantics

At the AIHPCcon Taiwan AI Supercomputing Conference held in May of 2023, TWSC unveiled the Formosa Foundation Model (FFM) series, featuring FFM-176B and FFM-7B, both trained on the open-source BLOOM large language model. Building upon the success of the FFM releases, FFM-Llama2 — which includes models 70B, 13B, and 7B — supports the development of the Code Llama program and allows for embedding models that enhance semantic search. FFM-Llama2 also offers various large language models for AI PC and provides models suitable for use in the cloud or on-premises. This flexibility, combined with comprehensive solutions offered by AFS, accelerates the application and realization of AI 2.0.

The traditional Chinese capabilities of FFM-Llama2 have been significantly enhanced through AIHPC supercomputing power, a high-performance parallel computing environment, advanced large language model division and partitioning technology, plus training with rich Taiwanese local language data. This versatile language model allows fine-tuning, deployment, and inference for all FFM-Llama2 models.

Enabling generative AI solutions

TWSC has recently released a new Parameter-Efficient Fine-Tuning (PEFT) feature for AFS. This feature offers in-the-cloud or on-premises inference options with a better price-performance ratio by optimizing parameters using GPTQ 4-bit high-performance compression technology. Since the release of FFM, AFS has been deployed to over 70 large enterprises. A notable example is Taiwan Mobile, a Taiwanese telecom company, which is implementing generative AI in its intelligent customer service system. This technology is expected to be extended to various industries to improve customer service efficiency.

By leveraging its AIHPC Supercomputer Service, Container Compute Service (CCS) with built-in generative AI image files, AFS solutions, and OneAI Machine Learning no-code AI platform, TWSC offers generative AI solutions that prioritize reliability, controllability, data security, and on-premises deployment.