China Telecom Artificial Intelligence Technology Co., Ltd. has made the Xing4.0-29B-A4B model available, which is the latest development in the Xing series (formerly known as TeleChat). The weights and configuration files for this model are hosted on Hugging Face and ModelScope under the XingChen-AGI organization. Officially, the model is positioned as China Telecom / XingChen / Xing4.0.
The model is presented as an agentic implementation of Mixture-of-Experts (MoE), optimized for the Ascend platform, rather than just a general release of a large language model. According to company materials, this is the first model of this scale fully trained on Ascend NPUs using the MindSpore framework and specifically tuned for complex engineering tasks and agent workloads.
According to Hugging Face data, the model has a total of 29 billion parameters, with approximately 4 billion parameters activated per token. It supports a native context length of 256K, which can be extended to 512K, features 40 layers, uses MLA attention, and includes 64 routed experts, of which four experts plus one shared expert are activated per token. Architectural notes point to an mHC + MLA + MTP stack designed for multi-step planning, tool calling, and maintaining coherence over long contexts.
It is reported that joint optimization of training on Ascend 910C clusters via MindSpore/MindFormers, including integrated mHC operators and MoE communication tuning, increased training throughput by approximately 96% compared to baseline settings, according to vendor data.
Various serving paths are provided for usage, such as Transformers, vLLM, SGLang, and KTransformers. Fine-tuning support is also available through LLaMA-Factory and MindFormers, along with formatting for agent systems, including OpenCode, Claude Code, OpenClaw, and Hermes. Furthermore, Phoenix Tech notes that coverage of secondary IT topics shows agent results comparable to leading Qwen models, and claims that 4-bit quantization can reduce memory consumption to 15GB for local long-context inference; however, these statements remain third-party or company claims pending independent verification.
For teams working on Huawei Ascend stacks, the Xing4.0-29B-A4B weights are provided with explicit training provenance on MindSpore and open interfaces for inference within the Apache ecosystem. Thus, this is an OSS agentic MoE from China Telecom, not a closed announcement based solely on an API.