LG open-sources a 750B model — K-EXAONE 2.0 under Apache 2.0
LG AI Research released K-EXAONE 2.0, a 750B-parameter MoE model, under Apache 2.0
The three lines
- K-EXAONE 2.0 ships under Apache 2.0 — commercial use, modification and redistribution allowed
- 750B total parameters with 37B active (MoE), and a 262,144-token context window
- It is the first frontier-scale fully open release from a Korean company
Key questions
- What makes K-EXAONE 2.0 notable
- The combination of scale and openness. A 750B-parameter mixture-of-experts model under Apache 2.0 — the license that permits commercial use outright — puts a Korean lab in territory previously held by a handful of players like DeepSeek. For Korea's industry it is a first at this scale.
- Why does it say 750B but only 37B active
- Mixture of Experts: the model holds 750B parameters of knowledge but wakes only the relevant ~37B 'expert' blocks per query, giving big-model knowledge at mid-model compute cost. It has become the standard architecture for large open releases — our reference piece today explains it.
- How good is it
- Self-reported benchmarks are top-tier for open models: 83.5 MMLU-Pro, 92.3 AIME 2026, 68.2 SWE-Bench Verified. But they are the vendor's own numbers — per our house rule, treat them as claims until independent verification lands.
The open-model map gained a Korean coordinate. LG AI Research released K-EXAONE 2.0 under Apache 2.0 — a 750B-parameter mixture-of-experts model that anyone may download, modify, and ship in commercial products.
1. What was released
The numbers: 750 billion total parameters, roughly 37 billion active per query, a 262,144-token context window (hundreds of book pages), support for ten languages including Korean and English. The license is the headline: Apache 2.0 is not a research-only arrangement — it is full commercial freedom.
Self-reported benchmarks — 83.5 on MMLU-Pro, 92.3 on AIME 2026, 68.2 on SWE-Bench Verified — would seat it among the strongest open models anywhere. The asterisk is standard: vendor-run numbers stay in the "claims" column until independent verification arrives.
2. Why it matters
| Item | Detail |
|---|---|
| Scale | 750B MoE (37B active) — top tier for open weights |
| License | Apache 2.0 — full commercial use |
| Context | 262,144 tokens |
| Languages | 10 |
| Benchmarks | 83.5 MMLU-Pro · 92.3 AIME · 68.2 SWE-Bench (self-run) |
| Significance | Korea's first frontier-scale fully open release |
Context: the open-weights race has been paced by China's DeepSeek while US big tech stayed mostly closed. A Korean conglomerate entering at top scale changes the map — and gives Korean companies and public institutions a domestic option that doesn't route through foreign APIs, a sovereignty argument that carries weight in Seoul.
3. What is still open
Verification and ecosystem: whether independent benchmarks confirm the claims, and whether cloud providers and enterprises actually deploy it. There is also a policy wrinkle — open weights sit awkwardly with the pre-release government testing framework Washington discussed this week (see today's companion pieces). For the vocabulary — open weights vs open source, how MoE works — today's reference "What 'open-source AI' actually means" has the map.