Upstage Launches Small Language Model on AWS - Amazon

To Help Businesses Around the World Build and Scale Generative AI Applications for the Korean Market
 

SEOUL - Amazon Web Services (AWS), an Amazon.com company, announced that leading South Korean artificial intelligence (AI) startup Upstage has launched its flagship SOLAR MINI Small Language Model (SLM) on the world’s leading cloud. SOLAR MINI is a versatile model that can be easily customized and fine-tuned to perform a wide range of language tasks in both Korean and English (Thai and Japanese will be available soon) such as understanding, summarizing, translating, and predicting new content. The model makes it easier for Upstage’s customers to get started with generative AI without having to train their models from scratch, and launch new applications that are tailored to industry-specific use cases, a press release stated by Amazon. 

Starting March 21, SOLAR MINI is available on Amazon SageMaker JumpStart, a machine learning (ML) hub offering foundation models, and on AWS Marketplace, AWS's curated digital catalog that makes it easy to deploy solutions from third-party software vendors. Upstage users can finetune SOLAR MINI leveraging domain-specific training data, which is already being used today by South Korean and global companies such as Metaverse Entertainment, Quora, ConnectWave, and Qanda.

“Upstage’s SOLAR MINI on AWS is a powerful solution that helps enterprises of all sizes to easily embrace generative AI at the local level,” said Jeongwon Yoon, country director, worldwide public sector, AWS Korea. “Language models are evolving rapidly and Upstage is at the forefront of this innovation, taking advantage of AWS’s secure and scalable stack to expand its AI offerings and provide organizations across industries with the tools they need to easily deploy generative AI to their applications.”

Language models enable generative AI services by providing context, memory capabilities, and text generation. Compared to large language models (LLMs) that are in the hundreds of billions of parameters, SLMs are lightweight and use less than 20 billion parameters. With a smaller training dataset of 10.7 billion parameters, SOLAR MINI can run inference (the process of using a trained machine learning (ML) model to make predictions based on new input data), at lower costs.

“From day one, we’ve designed Upstage to become a global AI player,” said Sung Kim, CEO of Upstage. “As Upstage’s preferred cloud provider, AWS has supported us in every step of this journey with programs like AWS Activate to access credits and technical support to rapidly scale our business. Today, we are thrilled to deepen our relationship with AWS to launch our SOLAR MINI model on Amazon SageMaker JumpStart, which opens up new possibilities for companies around the world to use the Korea’s most versatile LLM. Upstage is committed to redefining the AI landscape and delivering the most reliable and innovative AI solutions for corporate applications.”

Upstage selected AWS for its computing power and cost effectiveness to train SOLAR MINI, leveraging Amazon SageMaker, a service that enables developers to develop and deploy ML models.

Upstage’s advanced data pre-processing (steps taken to prepare and clean input data before it is used for training) and finetuning techniques, including Retrieval-Augmented Generation (RAG; the process of optimizing the output of a large language model), were used to train the model using data provided by the 1 Trillion Token Club, an industry-wide alliance founded by Upstage that is devoted to developing a large-scale language model that specializes in Korean. Consisting of a variety of high-quality, copyright-free Korean training data from texts, books, news articles, reports, and theses, the data evolved SOLAR MINI’s understanding of cultural nuances while ensuring higher accuracy in responses, preventing “AI hallucinations” that generate false or inappropriate answers.

In December 2023, SOLAR MINI was recognized as the top open-source large-scale SLM AI model by global AI platform HuggingFace’s Open LLM Leaderboard, scoring a 74.2 for its excellent design, encompassing the model’s reasoning, common sense inference, hallucination prevention, contextual understanding, and factual accuracy. This recognition demonstrates that SOLAR MINI performs as well as or better than bigger models, resulting in significant cost savings for graphics processing units (GPUs) that provide the horsepower to perform complex language tasks efficiently.

Accelerating generative AI innovation across all industries

E-commerce solutions provider ConnectWave has built a private LLM by training its proprietary data of 1.4 billion products with SOLAR MINI to automate e-commerce modernization, delivery tracking, inquiries, and return consultations for its price comparison websites Danawa and Enuri.

“By seamlessly integrating Upstage's advanced model with our dataset on AWS, we successfully crafted a customized, private LLM specifically designed for e-commerce applications,” said Kyung-sung Sohn, Senior Vice President of ConnectWave.

Education technology provider Qanda is working with Upstage to give students the tools they need to improve mathematics proficiency. Using SOLAR MINI, Qanda is developing a specialized language model that can solve problems and formulate descriptions and derivatives through mathematical reasoning. “Our partnership with Upstage is accelerating the development of state-of-the-art models and bringing newfound capabilities into the hands of our students,” said Min-gyu Cho, Head of Business of QANDA (Mathpresso).

In October 2023, AWS announced plans to invest KRW 7.85 trillion ($5.88 billion) in cloud infrastructure in South Korea by 2027 to meet the growing customer demand for cloud services. This investment is estimated to contribute KRW 15.06 trillion ($11.28 billion) to South Korea’s total gross domestic product (GDP) by 2027 and support an estimated average of 12,325 full-time equivalent (FTE) jobs in South Korean businesses each year.