Year: 2026
Type: preprint
Abstract: We present Xing 4, a series of large language models that extends the TeleChat lineage through coordinated advances in model architecture and training infrastructure on Ascend. This report describes the design principles and systems engineering behind the series, spanning architecture selection, large-scale pretraining, and reasoning and agentic post-training. We study how sparse activation, attention design, model depth and width, and speculative decoding balance model capacity with computation, memory access, and execution overhead. Our pretraining infrastructure builds on MindSpore HyperParallel, combining topology-aware hybrid sharding, efficient distributed optimizer execution, and coordinated management of long-sequence workloads and activation memory. CPU affinity and core isolation further reduce host scheduling interference. Together, these mechanisms support the pretraining of a 285B-parameter model on 8,192 Ascend 910C cards, providing a concrete setting for the architectural and systems studies. For post-training, we introduce TeleGym, a unified agent execution platform that connects execution, evaluation, and trajectory construction to policy optimization. Disaggregated asynchronous execution, bounded oversampling, and partial rollout accommodate heterogeneous task durations while preserving trajectory continuity across policy updates. Together, these advances establish an Ascend-oriented training foundation for the continued development of Xing 4 across model scales and increasingly demanding reasoning and agent workloads.
Source: arXiv (Cornell University)
Authors: Lab, Xingchen AGI, :, Hu, Baotian, Yang, Bingkai, Wang, Boyang, Pan, Changzai, Lyu, Chao, Wang, Chao, Li, Chen, Pan, Chenshuo, Lin, Chenyu, Zhang, Chenyuan, Liu, Chongming, Fan, Dawei, Wu, Di, Guo, Dong, Li, Dong, Li, Dongfang, Liu, Enlin, Xiao, Fanye, Li, Fengkai, Yao, Fubei, Lin, Gang, Gao, Han, Wu, Hanming, He, Haowei, Liu, Huapeng, Wang, Hua, Xu, Huinan, Xiong, Jiahong, Peng, Jiaxin, Ma, Jianxiang, Zhuge, Jianxing, Zhang, Jie, Xiong, Ji, Chen, Jinglin, Li, Jingqi, Li, Jingxuan, Ye, Jiwang, Li, Junchen, Li, Junbiao, Wu, Junyang, Yu, Kaidong, Jia, Kaipeng, Chen, Kehai, Li, Kuai, Liu, Laobin, Li, Lei, Qin, Libo, Shi, Lingling, Pu, Luwen, Xu, Menghao, Bao, Mengjiao, Chen, Mengqi, Zhang, Min, Wang, Mingrui, Wang, Naying, Han, Peicong, Guo, Peiming, Li, Pengfei, Feng, Qiang, Wang, Qi, Zhu, Qingjun, Li, Qingli, Ding, Qitai, Xie, Qiyi, Zhao, Quanchao, Han, Rongbao, Xue, Rui, Fang, Ruiyu, Song, Ruiting, Li, Shijie, Wang, Shiquan, Song, Shuangyong, Xiong, Sishi, Su, Teng, Bao, Wankang, Zhang, Weijie, Mao, Wei, Fu, Weiwei, Huang, Weiyang, Deng, Wenmin, Hu, Wenqing, Zhong, Wujie, Yu, Xi, Hu, Xiaohui, Fang, Xiaomin, Huang, Xiaomeng, Shi, Xiaoqiong, Shang, Xiaotong, Li, Xiaoyu, Xu, Xinglei, Jiang, Xin, Wang, Xin, Zhang, Xin, Liu, Xinzhang, Liu, Xue, Bai, Xuefeng, He, Xuewei, Li, Yahan, Zhang, Yanhan, Xue, Yanlei, Wu, Yanzhang, Yu, Yang, Liu, Yi, Zhang, Yin, Li, Yingyan, Yao, Yitong, Li, Yixuan, Xia, Yongqiang, Li, Yongxiang, Zhu, Yinzhou, Zhao, Yu, Jiang, Yuhao, Zhang, Yunpeng, Lin, Yuting, Wu, Yuxuan, Huang, Yuyao, Wang, Yuyi, Peng, Zebo, Pei, Zhenwu, Li, Zhenxiong, Wang, Zhao, Zhang, Zhaoxi, Yang, Zhihao, Dong, Zhiqiang, Yan, Zhiwei, Lu, Zhilong, Guo, Zhonghao, Wu, Zhongheng, He, Zhongjiang, Li, Zhongqiu, Yuan, Zhu, Jiang, Zhuoru, Wei, Zhuoyi, Guo, Zijia, Huang, Zilu, Guo, Zhipeng, Li, Zirui, Wang, Zihan
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: Harbin Institute of Technology
OpenAlex ID: W7236125976
