Vendor-funding-free attributes for Korean LLMs and the global models teams compare them against. Numbers are sourced; a paid diagnostic measures them on your prompts.
✦Independent · zero vendor funding·Data verified 2026-07-13·Methodology & sources ↗
Full comparison
| Model | Params | Context | License | Price in/out | Benchmark |
|---|---|---|---|---|---|
| Korean models | |||||
| Solar Pro 4 Upstage | 비공개 | 512K | API | $0.3 / $1.2 /1M | 업스테이지 발표: Terminal-Bench 2.1 57 |
| Solar Pro 3 Upstage | 확인되지 않음 | — | API | $0.15 / $0.6 /1M | Frontier LM Intelligence 리더보드 등재 (유일 한국 모델) |
| HyperCLOVA X (THINK) Naver | 비공개 (오픈웨이트 32B Think 별도 공개) | — | API | self-host / console | KoBALT-700 한국어 추론 상위 |
| EXAONE 4.0 32B LG AI Research | 32B (경량 1.2B 변형 존재) | — | terms unverified | self-host / console | Artificial Analysis 지능지수 62 (32B 최고) |
| A.X 4.0 SK Telecom | 72B / 7B Light | 131.072K | terms unverified | self-host / console | KMMLU 78.3 (GPT-4o 72.5 상회) |
| Trillion 7B Trillion Labs | 7.76B | 32.768K | open/commercial | self-host / console | KOBEST 0.795 (한국어 벤치 최상위) |
| Global comparison set | |||||
| Claude Anthropic | 비공개 | — | API | self-host / console | — |
| GPT OpenAI | 비공개 | — | API | self-host / console | — |
| Gemini | 비공개 | — | API | self-host / console | — |
Korean models — detail
Upstage publishes it for multilingual work including Korean, long context, and agents. It supports 512K context and up to 128K output; vendor performance claims still need independent workload testing.
Official sources checked 2026-08-23
Korean-optimized reasoning; the only Korean model on the Frontier intelligence leaderboard.
Strong Korean reasoning/search (high on Korean benchmarks); tight Korean data/search integration.
Robust EN+KO; highest Artificial Analysis intelligence (62) among 32B models.
KMMLU 78.3 (> GPT-4o 72.5); ~33% better Korean token efficiency than GPT-4o.
KOBEST 0.795 (top tier); strong Korean nuance for its size.
Global comparison set — detail
Top-tier Korean long-form instruction-following among global models.
Strong generalist; Korean cost-efficiency/honorific consistency can trail Korea-tuned models.
Long context + multimodal; Korean quality task-dependent.
Head-to-head
By use case
Try the models
Compare these models hands-on via OpenRouter. Affiliate link — a small share of usage supports KoreanLLM's independent operation and does not affect our evaluation.
Open in OpenRouter ↗Continue exploring