A little while ago, KT announced Mi:dm 2.0, its Korean-language AI LLM model. It was all over the news, which definitely reminded me that KT is a very large company.

Unfortunately, at the time, even if I searched Hugging Face, I could not find actual model information, and I could not even find a KT repository. SKT had one, but KT did not.

Then yesterday, almost on a whim, I checked again and suddenly there it was: KT’s Hugging Face page had appeared. K Intelligence Mi:dm had officially shown up.

So I decided to try it. It looks like Mi:dm 2.0 currently comes in two versions: a 2B-sized 2.0 Mini model and a 12B base model.

To run it easily on a MacBook, I used LM Studio.

Once I started testing it, what really stood out was how surprisingly fast it handled Korean. This is a local LLM, after all.

Maybe because it was trained on Korean writing and Korean information, it also answered the Dokdo question the way I expected.

Just from trying it briefly, it runs very quickly even on a local MacBook. It worked well not only on my M3 Max but also on a base-model M4 mini.

You can check the overall usage method and the actual demo in the video below.

Before trying Mi:dm 2.0, I honestly had some doubts about why a Korean-specialized AI was necessary at all, and why the concept of sovereign AI, which the Korean government also talks about a lot these days, was supposed to matter so much.

But after spending even a short amount of time with Mi:dm, I began to think there really is value in having a Korea-specific AI LLM that is less likely to produce nonsense when creating Korean-language content or educational materials such as history content for children.

Recently, a KAIST professor also explained why Korean AI matters by asking people to think about how foreign LLMs might answer a question like “Who does Dokdo belong to?”

After trying Mi:dm myself, I can understand that point a little better now.

Right.

So to close out this article, I decided to ask Mi:dm one more question:

“Who does Dokdo belong to?”

Huh?

Come on, you are not supposed to say that.

It looks like the road toward a truly solid Korean AI is still a long one.