LG AI Research has released K-EXAONE 2.0, a 750-billion-parameter mixture-of-experts model with 37 billion parameters active per token, under the Apache 2.0 licence. That licence is the news as much as the model: LG's previous K-EXAONE release carried a custom "k-exaone" licence with its own terms, and the company now writes in the model card that "we also release K-EXAONE 2.0 under the Apache license 2.0."
Key facts
- 750B total parameters, 37B active; 256 experts with 8 activated per token; 262,144-token context.
- Licence changed from a custom k-exaone licence on the previous 236B release to Apache 2.0.
- Supports 10 languages: Korean, English, Spanish, German, Japanese, Vietnamese, French, Italian, Polish and Portuguese.
- Primary source: the K-EXAONE 2.0 model card on Hugging Face, published July 29, 2026.
Licence text is the least glamorous part of a model release and often the only part that determines whether anyone can build on it. A custom licence, however reasonable its terms, forces every company's legal team to read it from scratch, and most will simply decline rather than spend the hours. Apache 2.0 is a known quantity that has been reviewed by essentially every legal department in the software industry. Moving from one to the other converts a model from something researchers evaluate into something products can ship.
That is a notable move for a national champion. LG is a Korean conglomerate, and the K-EXAONE line is explicitly a sovereign-AI effort - a flagship model built in Korea, for Korean, by a Korean company. The previous licence reflected that: a bespoke instrument giving LG control over how its work travelled. Choosing Apache 2.0 for the much larger successor trades that control for reach.
The scaling story is worth understanding because it is now the standard playbook. LG did not train 750 billion parameters from a blank slate. It describes upcycling the previous model to more than three times its size - taking a trained dense or smaller sparse network and expanding its feed-forward layers into many experts initialised from the original weights - then running continual pretraining, difficulty-focused mid-training and post-training on the enlarged model. The intuition is that you keep everything the smaller model already learned instead of paying for it again, and spend your compute teaching the new capacity to specialise. It is the same instinct as fine-tuning, applied to architecture rather than behaviour.
The model card also advertises two speculative decoding methods, multi-token prediction and one LG calls DSpark, which it says can speed generation by roughly three to five times. Speculative decoding works by having something cheap guess the next several tokens and having the full model verify them in one pass, so that a correct guess costs a fraction of what generating those tokens normally would. For a model this large, where each token is expensive, that is not a nicety.
On results, the picture is honest rather than triumphant. K-EXAONE 2.0 posts its strongest numbers where you would expect a long-context Korean flagship to be strong - near-ceiling scores on Korean long-context retrieval and on a Korean safety evaluation, and a very high score on a 2026 competition-mathematics set. On broad English reasoning and general-knowledge evaluations it is competitive with, rather than clearly ahead of, the leading open-weight models from Chinese and US labs. LG's own framing is measured: it calls the model "broadly competitive with leading open-weight models," which is a more careful sentence than most launch posts manage.
It arrives in a week of open-weight releases dense enough that individual launches are hard to hear over each other - Thinking Machines shipped Inkling-Small's 532 gigabytes of weights the same week, following Upstage's Solar Open 2, another Korean release, earlier in the month. Two 250-billion-plus open models from Korean labs in a single month is itself a data point about where this is being funded.
The caveat is the familiar one for models at this scale: 750 billion parameters have to be stored somewhere regardless of how few are active per token, so "open" here means auditable and fine-tunable rather than runnable on your desk. And a licence change on the model card is a statement of intent, not a warranty - anyone building commercially on it should read the actual LICENSE file in the repository rather than trusting the badge, which is good practice regardless of who published the model.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)