Upstage said on Oct. 1 that its Solar Pro 2 large language model posted the top score among Korean-built systems on a new government benchmark designed for enterprise document question-answering and compliance summarization. The National IT Industry Promotion Agency published the leaderboard alongside sample tasks drawn from anonymized public-sector forms.
Enterprise pricing
Upstage listed API pricing unchanged for October, betting benchmark publicity will convert free-tier developers before rivals cut rates. Chief executive Suk Hwan Lim told employees in an internal memo that public-sector pilots—not consumer chatbots—remain the firm’s margin engine for the rest of 2026.
The result gives Upstage a marketing win one day after Naver Cloud expanded commercial access to its HyperCLOVA Seed Think line. Enterprise buyers comparing vendors ahead of fourth-quarter budget cycles now have a third-party scorecard, not just vendor white papers.
What the benchmark tests
NIPA’s suite mixes Korean-language contract clauses, procurement regulations, and HR policy excerpts. Models must extract obligations, flag conflicting articles, and produce bullet summaries under token limits. Solar Pro 2 edged domestic rivals on accuracy metrics while matching open-weight imports on latency, according to the agency tables.
Officials stressed the benchmark does not certify models for regulated industries; financial and telecom buyers still need separate sandbox approvals from the Financial Supervisory Service and Korea Communications Commission.
Deployment paths
Upstage sells both API access and on-prem appliances for agencies that cannot send citizen data to public clouds. The Oct. 1 announcement highlighted a Gangwon provincial pilot that routes social-welfare inquiries through Solar Pro 2 with human review on every outbound message.
Competitors including NC AI and Naver are expected to submit updated weights before year-end. OpenAI and Anthropic models were included as references but were not eligible for the Korean-vendor category Upstage won.
Why enterprises care
Korean firms spent much of 2026 negotiating where inference may run after a series of payment-platform leaks. A government benchmark gives chief information officers cover to justify domestic models in board memos, even if production rollouts remain slow.
Upstage shares rose in early trading alongside other AI software names, though the Kospi’s chip-led rebound dwarfed those moves.
Caveats
Benchmarks age quickly; Solar Pro 2’s lead may narrow when multilingual tasks or code generation suites are added. NIPA said it will refresh tasks quarterly and publish failure cases, including hallucinated legal citations, to keep vendors honest.
For now, Oct. 1 belongs to Upstage in the domestic LLM score wars—a small headline in a week when parliament will spend far more airtime on submarines and housing permits than on model cards.