_version_ 1866910018374205440
author Cheon, Sung Jun
Cho, Jaekyung
Choi, Seongho
Eun, Hyunjun
Jo, Seokhwan
Jun, Jaehyun
Kang, Minsoo
Kim, Jin
Kim, Jiwon
Kim, Minsang
Kim, Seungsik
Kim, Sungwan
Kim, Tae Yoon
Kim, Youngrang
Lee, Hyeongmun
Lee, Sangyeol
Lee, Sungeun
Lee, Youngsoon
Lee, Yujin
Ok, Seongmin
Park, Chanyong
Park, Hyewoong
Park, Junyoung
Yang, Hyunho
Yi, Subin
Arya, Dhammiko
Bae, Soohyun
Cho, Dongyeon
Cho, Seungmo
Choi, Sangho
Choi, Yongseok
Han, Gyoungeun
Han, Yong-jin
Hong, Seokyoung
Hwang, Hyeon
Jang, Wonbeom
Ju, Minjeong
Jung, Wonjin
Ka, Keummin
Kang, Sungil
Kim, Dongnam
Kim, Jonghwi
Kim, Joonghoon
Kim, SaeRom
Kim, Sangjin
Kim, Seongwon
Kim, Youngjin
Lee, Seojin
Lee, Sunwoo
Lee, Taehoon
Park, Chanwoo
Park, Sohee
Park, Sooyeon
Ra, Yohan
Sek, Sereimony
Seo, Seungyeon
Song, Gun
Woo, Sanghoon
Yoon, Janghan
Yoon, Sungbin
author_facet Cheon, Sung Jun
Cho, Jaekyung
Choi, Seongho
Eun, Hyunjun
Jo, Seokhwan
Jun, Jaehyun
Kang, Minsoo
Kim, Jin
Kim, Jiwon
Kim, Minsang
Kim, Seungsik
Kim, Sungwan
Kim, Tae Yoon
Kim, Youngrang
Lee, Hyeongmun
Lee, Sangyeol
Lee, Sungeun
Lee, Youngsoon
Lee, Yujin
Ok, Seongmin
Park, Chanyong
Park, Hyewoong
Park, Junyoung
Yang, Hyunho
Yi, Subin
Arya, Dhammiko
Bae, Soohyun
Cho, Dongyeon
Cho, Seungmo
Choi, Sangho
Choi, Yongseok
Han, Gyoungeun
Han, Yong-jin
Hong, Seokyoung
Hwang, Hyeon
Jang, Wonbeom
Ju, Minjeong
Jung, Wonjin
Ka, Keummin
Kang, Sungil
Kim, Dongnam
Kim, Jonghwi
Kim, Joonghoon
Kim, SaeRom
Kim, Sangjin
Kim, Seongwon
Kim, Youngjin
Lee, Seojin
Lee, Sunwoo
Lee, Taehoon
Park, Chanwoo
Park, Sohee
Park, Sooyeon
Ra, Yohan
Sek, Sereimony
Seo, Seungyeon
Song, Gun
Woo, Sanghoon
Yoon, Janghan
Yoon, Sungbin
contents We introduce A.X K1, a 519B-parameter Mixture-of-Experts (MoE) language model trained from scratch. Our design leverages scaling laws to optimize training configurations and vocabulary size under fixed computational budgets. A.X K1 is pre-trained on a corpus of approximately 10T tokens, curated by a multi-stage data processing pipeline. Designed to bridge the gap between reasoning capability and inference efficiency, A.X K1 supports explicitly controllable reasoning to facilitate scalable deployment across diverse real-world scenarios. We propose a simple yet effective Think-Fusion training recipe, enabling user-controlled switching between thinking and non-thinking modes within a single unified model. Extensive evaluations demonstrate that A.X K1 achieves performance competitive with leading open-source models, while establishing a distinctive advantage in Korean-language benchmarks.
format Preprint
id arxiv_https___arxiv_org_abs_2601_09200
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle A.X K1 Technical Report
Cheon, Sung Jun
Cho, Jaekyung
Choi, Seongho
Eun, Hyunjun
Jo, Seokhwan
Jun, Jaehyun
Kang, Minsoo
Kim, Jin
Kim, Jiwon
Kim, Minsang
Kim, Seungsik
Kim, Sungwan
Kim, Tae Yoon
Kim, Youngrang
Lee, Hyeongmun
Lee, Sangyeol
Lee, Sungeun
Lee, Youngsoon
Lee, Yujin
Ok, Seongmin
Park, Chanyong
Park, Hyewoong
Park, Junyoung
Yang, Hyunho
Yi, Subin
Arya, Dhammiko
Bae, Soohyun
Cho, Dongyeon
Cho, Seungmo
Choi, Sangho
Choi, Yongseok
Han, Gyoungeun
Han, Yong-jin
Hong, Seokyoung
Hwang, Hyeon
Jang, Wonbeom
Ju, Minjeong
Jung, Wonjin
Ka, Keummin
Kang, Sungil
Kim, Dongnam
Kim, Jonghwi
Kim, Joonghoon
Kim, SaeRom
Kim, Sangjin
Kim, Seongwon
Kim, Youngjin
Lee, Seojin
Lee, Sunwoo
Lee, Taehoon
Park, Chanwoo
Park, Sohee
Park, Sooyeon
Ra, Yohan
Sek, Sereimony
Seo, Seungyeon
Song, Gun
Woo, Sanghoon
Yoon, Janghan
Yoon, Sungbin
Computation and Language
Artificial Intelligence
We introduce A.X K1, a 519B-parameter Mixture-of-Experts (MoE) language model trained from scratch. Our design leverages scaling laws to optimize training configurations and vocabulary size under fixed computational budgets. A.X K1 is pre-trained on a corpus of approximately 10T tokens, curated by a multi-stage data processing pipeline. Designed to bridge the gap between reasoning capability and inference efficiency, A.X K1 supports explicitly controllable reasoning to facilitate scalable deployment across diverse real-world scenarios. We propose a simple yet effective Think-Fusion training recipe, enabling user-controlled switching between thinking and non-thinking modes within a single unified model. Extensive evaluations demonstrate that A.X K1 achieves performance competitive with leading open-source models, while establishing a distinctive advantage in Korean-language benchmarks.
title A.X K1 Technical Report
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2601.09200