New Mixture-of-Experts Model Released by China Telecom

China Telecom Artificial Intelligence Technology Co., Ltd. has released Xing4.0-29B-A4B, a new large language model (LLM) within its Xing series, formerly known as TeleChat. According to documentation provided via the developer’s GitHub repository and shared on platforms such as Reddit, the model utilizes a Mixture-of-Experts (MoE) architecture.

The model features a total of 29 billion parameters, though it is designed to activate only 4 billion parameters per token. It supports a native context length of 256K tokens, which the developers state is extensible to 512K. Technical specifications indicate the model consists of 40 layers with a hidden size of 3584 and utilizes Multi-Head Latent Attention (MLA).

Hardware Optimization and Training Efficiency

The Xing4.0-29B-A4B is noted as the first model of its scale trained entirely on the Ascend NPU platform using the MindSpore framework. The developers claim significant training efficiency gains through multi-level co-optimization, including fine-grained MoE communication optimization and selective recomputation. According to the developer’s technical highlights, these optimizations resulted in an approximate 96% improvement in training throughput compared to out-of-the-box performance on Ascend 910C clusters.

The architecture is described as “agent-oriented,” built on an mHC + MLA + MTP framework. This design is intended to support multi-step planning, tool calling, and complex reasoning chains.

Benchmark Performance and Deployment

In comparative benchmarks shared by the developers, Xing4.0-29B-A4B demonstrated high performance in specific technical categories. According to the provided data, the model outperformed competitors such as Gemma4-26B-A4B and Qwen3.6-35B-A3B in several areas, including:

  • Claw-Eval: 76.55
  • SWE-bench Verified: 75.00
  • Terminal-Bench 2.1: 57.50
  • DeepresearchBII: 60.80

However, it trailed behind Qwen3.6-35B-A3B in the AIME2026 benchmark (90.00 vs 92.70) and IFBench (69.67 vs 72.67).

For deployment, the model maintains compatibility with several existing ecosystems, including LLaMA-Factory and MindFormers for fine-tuning, as well as vLLM and SGLang for inference. The developers suggest the model is optimized for vertical domain applications such as contract auditing, table understanding, and knowledge-based QA.

Community Observations

Discussions on Reddit regarding the release have noted the model’s weight, with one user observing that the 4-bit quantized version maintains a size of approximately 19GB. While the technical specifications are detailed, some community members have expressed interest in the release of the underlying training code to further verify the reported efficiency gains.