TeleChat to Xing4: Advancing Frontier Agentic Models on Domestic Chips and Frameworks

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.

Authors

Publication Details

Journal
arXiv (Cornell University)
Published
2026-11-17
DOI
https://doi.org/10.48550/arxiv.2611.22115
Primary Topic
Topic Modeling
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

TeleChat to Xing4: Advancing Frontier Agentic Models on Domestic Chips and Frameworks

Junyang Wu, Changzai Pan, Mengjiao Bao, Libo Qin et al.
arXiv (Cornell University)
Topic Modeling
preprint

TeleChat to Xing4: Advancing Frontier Agentic Models on Domestic Chips and Frameworks

Junyang Wu, Changzai Pan, Mengjiao Bao, Libo Qin, Chenshuo Pan, Xuewei He, Han Peicong, Luwen Pu, Peiming Guo, Rongbao Han, Quanchao Zhao, Chongming Liu, 刘恩林, Kuai Li, Menghao Xu, Chao Lyu, Hanming Wu, Bingkai Yang, Sishi Xiong, Ruiyu Fang, Qingli Li, Wankang Bao, Fubei Yao, Fengkai Li, Jingqi Li, Ruiting Song, Wenmin Deng, Qiyi Xie, Wenqing Hu, Xiaomin Fang, Huinan Xu, Teng Su, Weiyang Huang, Lingling Shi, Jianxing Zhuge, Jianxiang Ma, Xinglei Xu, Qitai Ding, Kaipeng Jia, Qingjun Zhu, Dong Li, Xiaoyu Li, Chen Li, Wujie Zhong, Xin Wang, Qiang Feng, Gang Lin, Xiaomeng Huang, Dong Guo, Junbiao Li, Fanye Xiao, Rui Xue, Xingchen AGI Lab, Jiahong Xiong, Haowei He, Weiwei Fu, Xiaoqiong Shi, Laobin Liu, Yahan Li, Jiwang Ye, Xiaotong Shang, Mengqi Chen, Xi Yu, Naying Wang, Kehai Chen, Shuangyong Song, Hua Wang, Xuefeng Bai, Shijie Li, Shiquan Wang, Chao Wang, Han Gao, Jingxuan Li, Weijie Zhang, Di Wu, Dongfang Li, Xiaohui Hu, Xin Zhang, Jinglin Chen, Wei Mao, Chenyuan Zhang, Qi Wang, Baotian Hu, Jie Zhang, Ji Xiong, Huapeng Liu, Lei Li, Jiaxin Peng, Boyang Wang, Pengfei Li, Xinzhang Liu, Dawei Fan, Xue Liu, Min Zhang, Xin Jiang, Junchen Li, Chenyu Lin, Mingrui Wang, Kaidong Yu
preprint en

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.

arXiv (Cornell University)
Topic Modeling
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon