China Telecom Releases MoE Model Based on Ascend NPU

China Telecom AI Technology Co., Ltd. has released its next-generation large language model Xing4.0-29B-A4B. The model adopts a MoE architecture that activates only 4B parameters per token out of a total of 29B, supports a base context length of 256K, and is extendable up to 512K.

Ascend NPU Optimization and Performance

This is the first model of its scale to be fully trained using the Ascend NPU platform and the MindSpore framework. By adapting the mHC and MLA architectures to the Ascend 910C cluster and developing fused operators, training throughput was improved by approximately 96% compared to baseline performance.

Agent-Oriented Architecture and Benchmarks

Based on the mHC, MLA, and MTP architectures, the model supports multi-step planning, tool calling, and execution of complex reasoning chains, ensuring task consistency in long contexts. In benchmark results, it scored 75.00 on SWE-bench Verified and 90.00 on AIME2026, demonstrating high performance compared to competing models.