Xing4.0-29B-A4B: China Telecom's Moe Model Trained on Ascend NPUs (www.mindstudio.ai)

China Telecom has unveiled the Xing4.0-29B-A4B, an innovative open-weight mixture-of-experts (MoE) language model designed for agentic tasks such as coding and multi-step planning. With 29 billion total parameters and an activating capacity of 4 billion per token, this model distinguishes itself by being the first of its scale fully trained on Huawei’s Ascend NPUs, utilizing the MindSpore and MindFormers software stack. This achievement highlights a significant shift as it deviates from the prevalent reliance on Nvidia GPUs, demonstrating that robust models can be developed on alternative hardware while achieving a reported 96% improvement in training throughput through advanced optimization techniques. The Xing4.0-29B-A4B excels in performance benchmarks tailored for terminal and agentic interactions, leading in metrics such as Terminal-Bench 2.1 and Claw-Eval compared to similarly sized models. The architecture supports a 256K context window, extendable to 512K, which is conducive for long-context agent tasks, augmented by innovative techniques like multi-token prediction and refined MoE communication. While competitive in agent-oriented benchmarks, it faces challenges in general reasoning tasks compared to rivals, making it an excellent candidate for teams focused on specialized coding and terminal automation workflows, all while offering an open Apache 2.0 license that facilitates broader commercial use.