SAN JOSE, Calif.— Dec 4, 2024— Leading South Korean artificial intelligence (AI) startup Upstage today announced that it has launched its next-generation large language model (LLM) Solar Pro, on Amazon Web Services (AWS). Available now on Amazon Bedrock Marketplace, Amazon SageMaker JumpStart and AWS Marketplace, Solar Pro can be easily customized and fine-tuned across a range of industries like healthcare, finance, and legal services.
At 22B parameters, Solar Pro, a larger model than Upstage’s previous 10.7B parameter model Solar Mini, shows 50% improvement in performance across key benchmarks at lower cost. In its September preview release, Solar Pro topped HuggingFace’s Open LLM Leaderboard for models under 70B parameters, excelled on the EQ Bench Leaderboard for emotional intelligence, and outperformed all other open-source models from major tech companies on Predibase's Fine-Tuning Leaderboard.
Solar Pro was trained on Amazon SageMaker, a fully managed machine learning (ML) service, that reduced training time significantly through advanced data pre-processing (steps taken to prepare and clean input data before it is used for training) and continued pre-training techniques. Data from the 1 Trillion Token Club, an Upstage-founded alliance for Korean-specialized LLMs, was used for training. Consisting of a variety of copyright-free English and Korean training data from texts, books, news articles, reports, the data evolved Solar Pro’s understanding of cultural nuances while ensuring higher accuracy in responses, mitigating “AI hallucinations” that generate false or inappropriate answers.
The 22B model is optimized for single-GPU deployment, utilizing Upstage's proprietary Depth-Up Scaling (DUS) method—a pre-training technique that ensures a compact model size without sacrificing performance. This method involves carefully redesigning their existing AI architecture to maintain high performance while significantly reducing its size. Without this technique, Upstage's AI model would need to be nearly four times larger to achieve the same results. By using DUS, Upstage created a 22B parameter model that performs on a par with much larger parameter model but requires far less computing power to run. Combined with specially selected training data, this approach helps create an AI model that can handle complex tasks efficiently, in turn, maximizing GPU efficiency, and allowing businesses to harness generative AI capabilities without expensive infrastructure upgrades or dependence on external APIs for data processing.
Starting today, Solar Pro is available in Amazon Bedrock Marketplace, where customers can now choose from more than 100 popular, emerging, and specialized models to find the optimal model for their use case. Once deployed, they can securely integrate the model with Amazon Bedrock’s unified application programming interfaces (APIs), leverage tools like Guardrails and Agents, and benefit from built-in security and privacy features. After a customer finds a model they want, they select the appropriate infrastructure (i.e., the number of instances and instance types) for their scaling needs, and easily deploy on AWS through fully managed endpoints.
AI Solutions Powering Industries
Solar Pro is able to intelligently comprehend document images and structured data, providing intelligent, context-aware responses to complex queries. The model excels in handling multi-page documents with varying layouts, automatically extracting and organizing critical information from invoices, contracts, and regulatory filings. For example, Solar Pro could process a structured multi-page invoice, regardless of its layout, and automatically extract key information such as invoice number, date, line items, total amount, and vendor details. It could then organize this information into a structured format and input it directly into the company's accounting system.
“Upstage's Solar Pro addresses critical challenges in industries like healthcare, finance, and legal, where automating and streamlining processes is essential for maximizing efficiency and productivity,” said Sung Kim, CEO of Upstage. “By harnessing AWS's powerful infrastructure, particularly Amazon SageMaker, we've developed a model that can interpret complex medical records, analyze financial reports, and process legal documents with high accuracy and efficiency, allowing professionals in these fields to make faster, more informed decisions. The scalability of AWS has enabled us to train Solar Pro on extensive datasets, reducing our training time and making advanced AI capabilities more accessible and cost-effective for businesses of all sizes.”
“AWS is proud to support Upstage in making advanced AI capabilities through Solar Pro more accessible,” said Kee Ho Ham, managing director of AWS Korea. “Our collaboration with startups like Upstage enables businesses to safely adopt and customize generative AI solutions that drive efficiencies. The AWS Partner Network (APN) has provided a structured framework for Upstage to build, market, and sell its generative AI solutions to organizations worldwide. Our technical resources, marketing support, and business development assistance are all designed specifically for startups to accelerate innovation across all industries, and Upstage is changing how custom AI solutions serve businesses in South Korea and beyond.”
Today, hundreds of global companies, including Intel, Poe by Quora, you.com, Sendbird are partnering with Upstage to leverage Solar Pro's capabilities, developing custom generative AI solutions tailored to their specific industry needs and challenges. This widespread adoption highlights Solar Pro's versatility and effectiveness across diverse sectors.
For example, South Korea's Ministry of Food & Drug Safety (MFDS) partnered with Upstage to develop an LLM-powered chatbot for public use. Trained on MFDS’s internal manuals, the chatbot provides real-time answers to questions about cosmetics regulations, including import, export, and sales requirements for products, streamlining a process that previously required MFDS officials to handle inquiries directly.
“The emergence of ChatGPT has brought significant societal changes, sparking growing interest in the public sector’s use of AI,” said Chanyoung Park, deputy director of MFDS. “At MFDS, we are actively exploring ways to leverage generative AI to streamline operations and enhance the delivery of food and drug safety information.”
Sendbird, a leading in-app communications solution provider that offers communication APIs and customizable AI chatbots, has integrated Upstage's Solar into its no-code generative AI chatbot solution that anyone can use to build a custom AI chatbot. This collaboration empowers businesses to rapidly deploy generative AI chatbots on websites and mobile apps, delivering human-like customer support and engagement powered by Solar LLM.
“We anticipate offering solutions that fully leverage the potential of generative AI by combining Sendbird AI Chatbot with Solar,” said John Kim, CEO of Sendbird. "It will become a game-changer for companies aiming for more innovative enterprise support services and customer experiences."
###
About Upstage
Founded in October 2020, Upstage boosts work efficiency with industry-leading document processing engines and large-language models (LLMs). Upstage’s flagship product Solar LLM delivers GPT-4-level performance with an unparalleled speed and cost-efficiency. Available via on-premises as well as API integration through platforms like Amazon SageMaker JumpStart, Solar Pro provides a versatile and accessible alternative to larger, more resource-intensive models developed by tech giants. Furthermore, Upstage’s Document AI solution leverages AI-powered optical character recognition (OCR) technology to automate workflows and process unstructured data, reducing operational costs and streamlining operations for our clients. Hundreds of customers worldwide leverage Upstage’s services including access to LLM models such as Solar Mini and Solar Pro through APIs and onboard/on-premises LLM development for customers including QUALCOMM, Intel, and Jasmine Technology Solution.