Generative AI (GenAI) and its innovations have disrupted industries worldwide, sparking a wave of transformation across sectors. As GenAI applications continue to expand, we are witnessing numerous industrial revolutions. Undoubtedly, NVIDIA is playing a pivotal role in this AI-driven era. Their technology evolves at an accelerated pace, and startups leveraging their products and solutions are rapidly advancing on their digital transformation journeys. APMIC, a Taiwan-based startup is one of the members of NVIDIA's Inception Program and also NVIDIA generative AI's ISV.

With NLU as the driving force for its development strategy, APMIC has created a large language model platform and trained a localized large language model (LLM) called CaiGunn. Enterprises using CaiGunn only need to upload their documents and articles or even embed their website URLs to start training their chatbot without writing a single line of code.

At COMPUTEX, APMIC announced the launch of their latest Traditional Chinese LLM called Llama-3-Taiwan-70B-Instruct, specially trained for the Taiwanese market. The model uses NeMo Retriever to perform retrieval augmented generation (RAG) on its 70 billion parameter model, resulting in chatbots that are more aligned with the Taiwanese culture, providing the necessary support.

The "Llama-3-Taiwan-70B-Instruct" model is the result of the efforts of many manufacturers. "APMIC is a reliable service provider for 'Llama-3-Taiwan-70B-Instruct', providing necessary solutions required by customers," NVIDIA Director Xiao Yiqi commented.

Project TAME (Taiwan Mixture of Experts), led by the NTU Department of Computer Science and Engineering, along with other researchers, to train a large language model (LLM) for Traditional Chinese as used in Taiwan, utilizing technology from the NVIDIA developer program was built on CaiGunn. This project was the champion of the SOTA (State-of-the-Art) in Traditional Chinese in Taiwan. During the Traditional Chinese test, its score even exceeded GPT4-turbo.

The product was trained using NVIDIA DGX H100s, as the 70B model is well-suited for the DGX and HGX H100 series hardware. Now, enterprises using APMIC's platform can directly select the Llama-3-Taiwan-70B-Instruct model to begin training with Traditional Chinese. Thanks to CaiGunn's no-code platform, nearly every employee in an enterprise can easily train their own AI.

With this development, Taiwanese enterprises can finally have their own 'AI assistant' that communicates in a familiar language. The model has been optimized by NVIDIA NIM to deliver improved inference speeds. It will soon support hardware environments from ASUS, Dell, HPE, Supermicro, and more, empowering each enterprise to rapidly deploy a dedicated intelligent system that drives operational excellence.

Join CaiGunn for free.

APMIC founder and CEO Jerry Wu highlighted that in the past, building a large language model (LLM) involved a series of complex processes. Starting with data input, business teams had to perform tasks such as converting PDFs, Word documents, and other file formats, storing the data in the file system, and converting text to NoSQL (non-relational databases). Then, the data format had to be placed into a vector database for LLM-based Retrieval-Augmented Generation (RAG) preparation. Following that, depending on the data and required accuracy, tasks like model fine-tuning, instruction fine-tuning, and Reinforced Learning with Human Feedback (RLHF) needed to be carried out. Finally, once the model version was obtained, users could upload data, provide prompts, and receive the final AI results. “Now, with APMIC’s ‘CaiGunn Talk,’ these processes can be easily automated. Combined with the Traditional Chinese outputs generated by Project TAME (Taiwan Mixture of Experts), we’ve created a ‘plug-and-play’ platform for large-scale language models. This platform enables natural language communication and allows every employee to harness AI to generate value for the company,” shared Jerry Wu.

Want to know more about CaiGunn and our LLM? Send us an email at support@ap-mic.com.