China Telecom Releases Xing4.0-29B-A4B, a 29B MoE Model Trained Entirely on Ascend NPUs

China Telecom Artificial Intelligence Technology Co., Ltd. has released Xing4.0-29B-A4B, a mixture-of-experts (MoE) large language model with 29 billion total parameters and 4 billion activated per token. The model, part of the Xing series (formerly known as TeleChat), natively supports a 256K token context window, extensible to 512K, and is available now on Hugging Face.

According to China Telecom, Xing4.0-29B-A4B is the first model of this scale trained entirely on the Huawei Ascend NPU platform using the MindSpore framework, rather than on Nvidia GPUs. The company says deep co-optimization with Ascend 910C clusters — including fine-grained MoE communication optimization, selective recomputation, DVM automatic graph-operator fusion, and custom Ascend C fused operators — delivered a training throughput improvement of approximately 96% over out-of-the-box performance.

Architecture

The model uses an architecture the company calls mHC + MLA + MTP, combining multi-head latent attention (MLA) with 64 routed experts, 4 active experts per token, and 1 shared expert. It has 40 layers, a hidden size of 3584, and is designed for agentic use cases: multi-step planning, tool calling, and long-context reasoning chains. Weights are released in F32 and BF16 formats and are compatible with Transformers, vLLM, SGLang, and KTransformers, as well as fine-tuning frameworks LLaMA-Factory and MindFormers.

Benchmarks

China Telecom published comparisons against Gemma4-26B-A4B and Qwen3.6-35B-A3B across nine benchmarks. Xing4.0-29B-A4B scored 75.00 on SWE-bench Verified (versus 53.00 for Gemma4-26B-A4B and 76.00 for Qwen3.6-35B-A3B), 57.50 on Terminal-Bench 2.1 (versus 30.00 and 51.50), and 90.00 on AIME2026 (versus 88.30 and 92.70). On Claw-Eval it posted 76.55, the highest of the three models listed. On IFBench it scored 69.67, trailing Gemma4-26B-A4B's 72.67. These figures come from China Telecom's own evaluation harnesses and have not been independently verified.

The model supports an OpenAI-compatible API and includes a toggleable "thinking" mode via chat template parameters. Recommended inference settings differ by task: temperature 1.0 for complex reasoning, temperature 0.8 for coding and agent workloads. The company has also aligned the model's format with agent frameworks including OpenCode, Claude Code, OpenClaw, and Hermes. Pricing for hosted access has not been disclosed, and no inference provider currently deploys the model.

What this means

Xing4.0-29B-A4B is notable less for raw benchmark leadership — it trails Qwen3.6-35B-A3B on several coding and reasoning tasks — and more for its training infrastructure. A model of this scale trained entirely on Ascend NPUs with MindSpore, rather than Nvidia hardware, signals continued progress in China's push toward domestic AI compute independence amid export controls. The 96% throughput improvement claim, if it holds up under independent scrutiny, would be a meaningful data point for the viability of non-Nvidia training stacks at this parameter scale. For now, the model's real-world utility will depend on third-party evaluation and whether it becomes available through hosted inference providers, since none currently list it.