← Latest papers
💬 NLP

A.X K1 Technical Report

A.X K1 is a 519B-parameter Mixture-of-Experts (MoE) language model trained on 10T tokens that features a unique "Think-Fusion" training recipe, allowing users to switch between reasoning and non-reasoning modes for efficient, controllable deployment.

Original authors: Sung Jun Cheon, Jaekyung Cho, Seongho Choi, Hyunjun Eun, Seokhwan Jo, Jaehyun Jun, Minsoo Kang, Jin Kim, Jiwon Kim, Minsang Kim, Seungsik Kim, Sungwan Kim, Tae Yoon Kim, Youngrang Kim, Hyeongmun Lee
Published 2026-02-12
📖 3 min read☕ Coffee break read

Original authors: Sung Jun Cheon, Jaekyung Cho, Seongho Choi, Hyunjun Eun, Seokhwan Jo, Jaehyun Jun, Minsoo Kang, Jin Kim, Jiwon Kim, Minsang Kim, Seungsik Kim, Sungwan Kim, Tae Yoon Kim, Youngrang Kim, Hyeongmun Lee, Sangyeol Lee, Sungeun Lee, Youngsoon Lee, Yujin Lee, Seongmin Ok, Chanyong Park, Hyewoong Park, Junyoung Park, Hyunho Yang, Subin Yi, Dhammiko Arya, Soohyun Bae, Dongyeon Cho, Seungmo Cho, Sangho Choi, Yongseok Choi, Gyoungeun Han, Yong-jin Han, Seokyoung Hong, Hyeon Hwang, Wonbeom Jang, Minjeong Ju, Wonjin Jung, Keummin Ka, Sungil Kang, Dongnam Kim, Jonghwi Kim, Joonghoon Kim, SaeRom Kim, Sangjin Kim, Seongwon Kim, Youngjin Kim, Seojin Lee, Sunwoo Lee, Taehoon Lee, Chanwoo Park, Sohee Park, Sooyeon Park, Yohan Ra, Sereimony Sek, Seungyeon Seo, Gun Song, Sanghoon Woo, Janghan Yoon, Sungbin Yoon

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Story of A.X K1: The "Smart Switch" Brain

Imagine you have a personal assistant. Most assistants are either "The Speedster" (who answers instantly but sometimes misses the nuances of a complex math problem) or "The Professor" (who is brilliant at solving deep mysteries but takes ten minutes to explain how to boil an egg).

Usually, you have to hire two different people. But SK Telecom has just introduced A.X K1, a massive new AI model that is essentially one person who can flip a switch between being a Speedster and a Professor whenever you need them.

Here is how they built this "super-assistant" using three clever tricks:


1. The "Expert Team" Architecture (Mixture-of-Experts)

Instead of building one giant, heavy brain where every single neuron has to fire for every single question, the engineers built a Mixture-of-Experts (MoE).

The Analogy: Imagine a massive hospital. If you walk in with a broken toe, you don't need a brain surgeon, a cardiologist, and a pediatrician all standing around your bed. That would be a waste of time and money. Instead, you have a receptionist who directs you to the specific specialist you need.

A.X K1 has 519 billion "knowledge pieces," but for any single question, it only "wakes up" about 33 billion of them. This makes the model incredibly smart (it has a huge library of knowledge) but very efficient (it doesn't waste energy using the whole library to answer "What is 2+2?").

2. The "Think-Fusion" Switch (The Best of Both Worlds)

This is the most unique part of the paper. Usually, "Reasoning Models" (like those that solve hard math) are forced to "think out loud" for every single prompt, which is slow and annoying for simple tasks.

The Analogy: Think of a professional chef.

  • Non-Thinking Mode: When you ask for a glass of water, they just hand it to you. Fast and direct.
  • Thinking Mode: When you ask for a 5-course wedding banquet, they sit down, grab a notepad, plan the menu, check the pantry, and then start cooking.

A.X K1 uses a special training recipe called "Think-Fusion." It allows the user to say, "Hey, just give me the quick answer," or "Take your time and think this through step-by-step." It can do both within the same "brain," preventing it from being a "slow professor" when you just need a "fast speedster."

3. The "Sovereign AI" Mission (The Cultural Specialist)

Most of the world's biggest AI models are trained primarily on English data from the US and Europe. This means they sometimes struggle with the specific "vibe," culture, and complex nuances of other languages.

The Analogy: It’s like a textbook written by someone who has never visited Korea. They might know the facts, but they don't understand the local slang, the social etiquette, or the subtle way people communicate.

A.X K1 was built to be a "Sovereign AI." SK Telecom poured massive amounts of high-quality Korean data into it. As a result, when it comes to Korean language tests, it doesn't just "translate" English logic—it actually understands the Korean context better than many of its global competitors.


Summary: Why does this matter?

In short, A.X K1 is a massive leap toward practical AI. It’s not just a "smart" model; it’s a flexible model. It’s designed to be smart enough to win math competitions, fast enough to power a chatbot, and culturally aware enough to feel like a local, all while being efficient enough to actually run on real-world computers.

Drowning in papers in your field?

Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.

Try Digest →