China Telecom Releases Xing4.0, an Open 29B Coding Agent Built on Huawei Chips
China Telecom's Xing4.0-29B-A4B is a 29B-parameter sparse MoE with 4B active, native 256K context, trained entirely on Huawei Ascend NPUs.
- China Telecom released Xing4.0-29B-A4B, a 29B MoE with 4B active parameters, Apache 2.0.
- Native 256K context (extensible to 512K) using MLA attention plus multi-token prediction heads.
- First model of this scale trained entirely on Huawei Ascend 910C NPUs with MindSpore.
- Hits 75 on SWE-bench Verified and 57.5 on Terminal-Bench 2.1, beating Gemma4-26B-A4B.
- Training throughput improved ~96% via fused operators and MoE communication tuning on Ascend.
- Ships with vLLM, SGLang, KTransformers support and adapters for Claude Code and OpenCode.
China Telecom releases a 29B MoE trained on Ascend
China Telecom’s AI unit has released Xing4.0-29B-A4B, an open-weight mixture-of-experts model aimed at coding agents, terminal work, and long-context tasks. The model contains 29 billion parameters, activates about 4 billion per token, supports a native 256K context window, and was trained end to end on Huawei Ascend NPUs.
Developers can download the weights under Apache 2.0 and deploy them through supported stacks including vLLM, SGLang, KTransformers, Transformers, and MindFormers. The release succeeds the TeleChat series and adds a production-scale example of MoE training on hardware outside NVIDIA’s ecosystem.
The release at a glance
| Developer | China Telecom Artificial Intelligence Technology Co., Ltd. |
|---|---|
| Parameters | 29B total, about 4B active per token |
| Architecture | 40-layer MoE with MLA, MTP, and an mHC block |
| Experts | 64 routed experts, 4 selected per token, plus 1 shared expert |
| Context | 256K native; the model card describes extension to 512K |
| Training stack | Huawei Ascend 910C, MindSpore, and MindFormers |
| License | Apache 2.0 |
| Primary workloads | Coding agents, terminal tasks, tool use, and long-context reasoning |
29B weights, 4B active
Xing4.0 uses sparse routing to send each token through four of its 64 routed experts and one shared expert. This reduces arithmetic per token while preserving a larger pool of learned parameters than a dense 4B model.
Total parameter count determines weight storage, while active parameter count mainly tracks per-token compute. At BF16 precision, 29 billion parameters require roughly 58GB for weights alone. Runtime overhead, activations, and the key-value cache add to that requirement, especially near the 256K context limit. An 80GB accelerator may accommodate shorter-context workloads, but capacity planning should use the full model size rather than the 4B active count.
- 40 transformer layers with a hidden size of 3,584.
- Dense feed-forward width of 9,216 and an expert intermediate size of 1,024.
- Multi-head Latent Attention
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
