_version_ 1866917224533458944
author K2 Team
Liu, Zhengzhong
Tang, Liping
Jin, Linghao
Li, Haonan
Ranjan, Nikhil
Fan, Desai
Rohatgi, Shaurya
Fan, Richard
Pangarkar, Omkar
Wang, Huijuan
Cheng, Zhoujun
Sun, Suqi
Han, Seungwook
Tan, Bowen
Gosal, Gurpreet
Han, Xudong
Pimpalkhute, Varad
Hao, Shibo
Hee, Ming Shan
Hestness, Joel
Jia, Haolong
Ma, Liqun
Singh, Aaryamonvikram
Soboleva, Daria
Vassilieva, Natalia
Wang, Renxi
Wu, Yingquan
Sun, Yuekai
Killian, Taylor
Moreno, Alexander
Maggs, John
Ren, Hector
He, Guowei
Wang, Hongyi
Ma, Xuezhe
Wang, Yuqi
Yurochkin, Mikhail
Xing, Eric P.
author_facet K2 Team
Liu, Zhengzhong
Tang, Liping
Jin, Linghao
Li, Haonan
Ranjan, Nikhil
Fan, Desai
Rohatgi, Shaurya
Fan, Richard
Pangarkar, Omkar
Wang, Huijuan
Cheng, Zhoujun
Sun, Suqi
Han, Seungwook
Tan, Bowen
Gosal, Gurpreet
Han, Xudong
Pimpalkhute, Varad
Hao, Shibo
Hee, Ming Shan
Hestness, Joel
Jia, Haolong
Ma, Liqun
Singh, Aaryamonvikram
Soboleva, Daria
Vassilieva, Natalia
Wang, Renxi
Wu, Yingquan
Sun, Yuekai
Killian, Taylor
Moreno, Alexander
Maggs, John
Ren, Hector
He, Guowei
Wang, Hongyi
Ma, Xuezhe
Wang, Yuqi
Yurochkin, Mikhail
Xing, Eric P.
contents We introduce K2-V2, a 360-open LLM built from scratch as a superior base for reasoning adaptation, in addition to functions such as conversation and knowledge retrieval from general LLMs. It stands as the strongest fully open model, rivals open-weight leaders in its size class, outperforms Qwen2.5-72B and approaches the performance of Qwen3-235B. We actively infuse domain knowledge, reasoning, long-context, and tool use throughout the training process. This explicitly prepares the model for complex reasoning tasks. We demonstrate this potential using simple supervised fine-tuning, establishing a strong baseline that indicates significant headroom for advanced alignment. By releasing the full training history and data composition, we maximize the effectiveness of continuous training, a key open source production scenario. We release the model weights and signature LLM360 artifacts, such as complete training data, to empower the community with a capable, reasoning-centric foundation.
format Preprint
id arxiv_https___arxiv_org_abs_2512_06201
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle K2-V2: A 360-Open, Reasoning-Enhanced LLM
K2 Team
Liu, Zhengzhong
Tang, Liping
Jin, Linghao
Li, Haonan
Ranjan, Nikhil
Fan, Desai
Rohatgi, Shaurya
Fan, Richard
Pangarkar, Omkar
Wang, Huijuan
Cheng, Zhoujun
Sun, Suqi
Han, Seungwook
Tan, Bowen
Gosal, Gurpreet
Han, Xudong
Pimpalkhute, Varad
Hao, Shibo
Hee, Ming Shan
Hestness, Joel
Jia, Haolong
Ma, Liqun
Singh, Aaryamonvikram
Soboleva, Daria
Vassilieva, Natalia
Wang, Renxi
Wu, Yingquan
Sun, Yuekai
Killian, Taylor
Moreno, Alexander
Maggs, John
Ren, Hector
He, Guowei
Wang, Hongyi
Ma, Xuezhe
Wang, Yuqi
Yurochkin, Mikhail
Xing, Eric P.
Machine Learning
We introduce K2-V2, a 360-open LLM built from scratch as a superior base for reasoning adaptation, in addition to functions such as conversation and knowledge retrieval from general LLMs. It stands as the strongest fully open model, rivals open-weight leaders in its size class, outperforms Qwen2.5-72B and approaches the performance of Qwen3-235B. We actively infuse domain knowledge, reasoning, long-context, and tool use throughout the training process. This explicitly prepares the model for complex reasoning tasks. We demonstrate this potential using simple supervised fine-tuning, establishing a strong baseline that indicates significant headroom for advanced alignment. By releasing the full training history and data composition, we maximize the effectiveness of continuous training, a key open source production scenario. We release the model weights and signature LLM360 artifacts, such as complete training data, to empower the community with a capable, reasoning-centric foundation.
title K2-V2: A 360-Open, Reasoning-Enhanced LLM
topic Machine Learning
url https://arxiv.org/abs/2512.06201