Upstage vs.
the giants.

On a quiet Tuesday in July 2025, the analyst firm Artificial Analysis updated its public leaderboard. Buried among the usual suspects - OpenAI, Anthropic, Google, Meta - a new name appeared in the "Frontier Model" tier. The name was Solar Pro 2. The company behind it, Upstage, has a US address on North 1st Street in San Jose and a much older one in Seoul. Until that update, almost no one outside Korea had heard of it. Now its 31-billion-parameter model was sitting five points above GPT-4.1 on the Intelligence Index, and the AI industry was suddenly paying attention to a Korean startup with roughly 150 employees.

If that sounds like a Hollywood arc, the reality is more grinding. Upstage didn't arrive at the frontier by accident. It got there by picking a fight that nobody else seemed interested in - building enterprise-grade AI that big regulated companies could actually deploy - and refusing to let go for five straight years.

Chapter OneThe problem they saw

Enterprise AI has a paperwork problem. Insurance carriers run on PDFs. Banks run on PDFs. Hospitals run on PDFs, prescriptions scribbled in three different scripts, and the kind of multi-page forms that make grown adults reach for coffee. Every large language model demo in the past three years has politely ignored this fact, opting instead for poetry and trivia. The companies that actually have money to spend on AI, meanwhile, have entire departments whose job is to retype scanned documents into spreadsheets.

The founders of Upstage - Sung Kim, Hwalsuk Lee, and Lucy Park - had spent years at Naver Clova, the AI lab inside Korea's largest internet company, watching this exact friction play out at scale. Sung Kim had been a computer-science professor at Hong Kong University of Science and Technology before that, the kind of academic who spends more time deploying things than publishing about them. Hwalsuk Lee led Naver's visual-AI team, which mostly meant OCR. Lucy Park led modeling on Papago, Naver's translation product. Three people whose entire professional lives had been spent watching messy documents and language collide.

They left in 2020 and started Upstage with what looked, at the time, like a niche thesis: the next decade of enterprise AI was going to be won by whoever could do the boring parts well.

Chapter TwoThe founders' bet

There is, of course, a particular brand of irony in three machine-learning researchers deciding their differentiator would be reading bad PDFs. But that's exactly what they bet on. The early Upstage product wasn't a chatbot. It was Document AI - an OCR-and-extraction engine that ingested forms, contracts, and scans and produced clean, structured JSON on the other end. Tables intact. Charts read. Korean and Japanese and English, all in the same document if that's what you needed.

For about two years they were that company. Korean insurers signed on. Then a few banks. Then, somewhat improbably, Samsung. In 2023, with the LLM gold rush in full swing, Upstage made a second bet that should have been a distraction and instead turned out to be a fulcrum: they built their own language model. They called it Solar.

Solar's distinguishing feature, then and now, is its size. The first version landed at 10.7 billion parameters - small enough to run on a single GPU, large enough to be useful. The thesis was simple and slightly contrarian: enterprises with regulated data don't want to send their documents to OpenAI. They want models they can deploy in their own data centers, on their own hardware, behind their own firewalls. Smaller models with sharper benchmarks would win those deals.

That bet aged well. Solar Pro 2, released in July 2025, weighs in at 31 billion parameters - still less than half the size of the frontier-class behemoths it now competes with. And yet on the Artificial Analysis Intelligence Index, it posts a score of 58. GPT-4.1 sits at 53. The benchmark community took a moment to recheck the numbers. They held.

Chapter ThreeThe milestones

Chapter FourThe proof

Benchmarks are pretty. Revenue is real. The case Upstage has been quietly assembling looks something like this: Samsung is using their document intelligence. A majority of the largest Korean insurers run claims through their models. Korean banks lean on them for KYC document processing. AWS lists their Document AI on its Marketplace and, more recently, on the AWS AI Agent Marketplace - the new shelf where Amazon expects enterprise buyers to actually go shopping for AI agents.

Amazon also bought equity in Upstage in 2025, as did AMD. That is unusual. Neither company has a habit of investing for the cocktail-party value. Amazon wants Solar on Bedrock-adjacent rails for enterprise customers who want a model that can read a 300-page insurance contract without losing track of clause 14.2(b). AMD wants Solar trained on AMD silicon. The Series B bridge round of $45M pushed total funding past $157M and put two of the largest names in cloud and compute on the same cap table.

Chapter FiveThe mission

Ask the founders what Upstage is for and you get an answer that is shorter than you'd expect. They want to make AI beneficial. Not transformative, not revolutionary - beneficial. The phrasing is deliberate, and slightly unfashionable in a market that prefers manifestos. The implied argument is that most AI hype is allergic to the small print, the regulated industries, the parts of the economy where AI has to be both good and provably good. That gap is where Upstage lives.

There is a culture clue in how they ship. The company publishes openly on Hugging Face. It competes on public leaderboards. It loses some of those competitions, and it talks about them anyway. There is something refreshing about a frontier-model lab that still acts a little like a research group that hasn't quite caught on to its own importance.

Chapter SixWhy it matters tomorrow

The conventional take on enterprise AI in 2026 is that the foundation-model layer will be a winner-take-most market dominated by three or four American labs. Upstage's counter-take is that this is true for chatbots and false for everything else. The places where AI will actually get deployed - regulated, on-prem, multilingual, paperwork-heavy - are the places where size, sovereignty, and document literacy matter more than benchmark sheen.

If that read is right, the company most American observers haven't heard of is going to keep showing up in the corners of the market that have the highest dollar density and the lowest media coverage. The deals will be quiet. The growth will be ugly-pretty - cohorts of insurance companies, hospital networks, manufacturers. The kind of revenue you can't tweet about.

Which brings us back to that Tuesday in July, and the analyst firm that quietly added a fifth name to its frontier list. The interesting thing about that update wasn't Solar Pro 2's score. The interesting thing was who put it there, and where they were from, and what they planned to do with it. Five years ago, three people walked out of Naver to read PDFs for a living. The PDFs are still being read. The model that reads them now happens to be one of the best in the world.

The frontier moved. It just didn't move in the direction most people were looking.