Upstage trains AI models faster and more reliably using OCI

Upstage is a Korean AI startup that provides AI-powered document processing and large language models that are particularly adept in Korean, Japanese, and other Asian languages, as well as English. While other AI models might be able to translate documents in these languages, Upstage’s models are built and trained to capture linguistic and cultural nuances.

“Other models are grammatically OK, but they sound translated,” Upstage CEO and Co-Founder Sung Kim says. To continuously train its AI models and develop new AI services in languages including Thai and Vietnamese, Upstage turned to Oracle Cloud Infrastructure (OCI) for computing power. With OCI, Upstage achieved approximately 30% faster data transfer within its GPU clusters, along with economical pricing and more reliable performance, all of which helped Upstage to bring AI models to market faster.

OCI gives Upstage speed, cost savings

The company’s mission is to “build intelligence for the future of work” by developing AI models trained to excel in select languages. Upstage chose OCI for training these AI models primarily because of OCI’s superior performance in three key attributes: speed, price, and reliability. Upstage runs a large number of GPUs at a time to train a model, and the company continuously trains and adapts models for new languages and business use cases. It needed infrastructure that could help it economically train models and bring them to market quickly.

Upstage also valued Oracle’s strong technical support and the opportunities for research collaboration. Upstage teams can work with Oracle AI and infrastructure experts to explore ways to use OCI most effectively and to train models more efficiently.

Upstage growth depends on speed to market

By providing AI models trained on OCI, Upstage landed more than 50 enterprise customers—including companies in financial services, healthcare, manufacturing, and law—for its large language model called Solar. It also saw success with its Document Parse service, which turns complex documents into formats that LLMs can easily process, such as pulling data from charts or scanned images. In addition, its AI-powered service can take data from documents with irregular formats, such as invoices coming from an array of suppliers. For all these use cases, Upstage used OCI GPUs to harness and clean the data needed for training, to perform the core AI model training, and then to fine-tune the model.

As a startup, Upstage needed to get to market fast with its AI models, and OCI helped it train models quickly. For example, Upstage engineers estimate that OCI’s data transfer—the first stage of Upstage’s AI model development—is at least 30% faster than that of other cloud providers the company has tried. When it comes to training models, reliability is also an underrated trait, CEO Kim says. If a model fails during training, the process must start over from the last checkpoint in the training cycle, which can result in losing two or three days of training work. Kim said that other providers’ GPUs would often fail two to three times a month, whereas OCI failed at most once and often not at all in a month. Upstage took two to three months to fully train its early AI models, so if it lost several days multiple times during a training, the company would have seen a significant slowdown. OCI proved highly reliable, a trait that’s just as important as OCI’s low cost for training. “We don’t care about a low price if the system isn’t stable,” Kim says.

View more Oracle Connect articles