Copied


Xing 4: Advancing Large Language Models on Domestic Ascend Chips

Caroline Bishop   Nov 18, 2026 06:36 0 Min Read


Xingchen AGI Lab has introduced Xing 4, a groundbreaking series of large language models (LLMs) that extend the TeleChat lineage. The research, published on November 17, 2026, on arXiv (2611.22115), highlights significant advancements in model architecture and training infrastructure, specifically tailored for domestic Ascend 910C chips.

Key Technical Innovations

Xing 4 is built on a 285-billion parameter architecture and trained using MindSpore HyperParallel, a framework optimized for large-scale distributed training. The infrastructure incorporates topology-aware hybrid sharding, efficient distributed optimizer execution, and CPU affinity with core isolation to minimize scheduling interference. These optimizations enable the use of 8,192 Ascend 910C cards, achieving efficient pretraining for long-sequence workloads and activation memory management.

The model architecture balances sparse activation, attention design, and speculative decoding to optimize computational and memory efficiency. A mixture-of-experts (MoE) topology, combined with lightweight correction heads, enhances speculative decoding performance. This approach supports a wide range of reasoning and agentic tasks while maintaining execution stability under long-context scenarios.

Post-Training Advancements

For post-training, the researchers developed TeleGym, a unified agent execution platform. TeleGym integrates execution, evaluation, and trajectory construction with policy optimization. Features such as disaggregated asynchronous execution, bounded oversampling, and partial rollout allow the platform to handle heterogeneous task durations while maintaining trajectory continuity across policy updates.

Implications for Domestic AI Development

Xing 4 represents a significant step forward in leveraging domestic hardware for advanced AI applications. By optimizing for Ascend NPUs, the research provides a scalable and efficient foundation for future LLM development, particularly in scenarios requiring complex reasoning and multi-step planning. The open-sourcing of a 29-billion parameter MoE member further underscores the project's commitment to advancing the AI community.


Read More