China Telecom Xing4.0-29B-A4B: 29B MoE With 4B Active Parameters and 256K Context

By AiCybr Editorial · Published September 22, 2026

China Telecom Artificial Intelligence Technology has released Xing4.0-29B-A4B, an Apache-2.0 model with 29 billion total parameters and about 4 billion activated per token. The sparse mixture-of-experts model supports a 256K-token native context window, with the project documenting an extension path to 512K, and targets coding agents, tool use and long-running engineering tasks.

The weights and deployment material became public on September 17. China Telecom AI says the model was trained entirely on Huawei Ascend NPUs with MindSpore and has also published FP8 and GGUF variants. The release is relevant to local-AI users because the vendor positions its quantized builds for single-GPU deployment while retaining a long context window and agent-oriented tool calling.

Key specifications

The architecture combines multi-head latent attention with multi-token prediction and a mixture-of-experts feed-forward design. China Telecom reports approximately 96% higher training throughput after its Ascend-specific communication, recomputation, graph-fusion and fused-operator optimizations compared with its unoptimized baseline. The 96% figure is a vendor engineering measurement from the model documentation.

Local deployment and current runtime support

China Telecom's September 22 release says a low-bit quantized build requires about 15 GB of GPU memory, enabling deployment on a single consumer graphics card. Total memory use also depends on KV cache, runtime buffers and context length, so deployments targeting the full 256K context need separate memory validation for the selected runtime and quantization.

The project's current README lists upstream support pull requests for vLLM, SGLang, TensorRT-LLM, llama.cpp and KTransformers as pending. Its serving instructions point users to the corresponding PR branches until those changes land upstream. The weights are publicly inspectable today while those upstream integrations continue.

The project provides a Transformers example and serving recipes for vLLM, SGLang and KTransformers. It also documents fine-tuning paths through LLaMA-Factory and MindFormers and targeted format adaptation for OpenCode, Claude Code, OpenClaw and Hermes.

Coding and agent benchmarks

China Telecom reports the following results from its evaluation suite:

These are vendor-reported benchmark results from the Xing4.0 model card. The model card also publishes harness details for several tests, including a 210K context for SWE-bench Verified and three-run averaging for Terminal-Bench 2.1. Deployment evaluation should use the exact quantization, agent harness, repository mix and hardware intended for production.

Why the 29B/4B design matters

The 4B active figure describes the subset of expert parameters used for each token; the deployment still stores the model's full expert weights. The practical attraction is compute efficiency from sparse activation combined with a model small enough to have quantized single-GPU options.

The release also provides a useful test case for an increasingly diversified AI training stack. China Telecom says Xing4.0 was trained entirely on Ascend hardware and MindSpore, then exposed through an ecosystem that also targets widely used GPU inference frameworks. That gives developers a concrete public artifact for evaluating portability across the two stacks.

For local deployment today, choose the published quantization first, budget memory for the desired context separately, and verify runtime-support status before building an agent workflow around it. The Apache-2.0 license and public weights permit direct local evaluation without a hosted API dependency.

Sources

• Xing4.0-29B-A4B official GitHub repository and model documentation
• Xing4.0-29B-A4B official Hugging Face model and weights
• Xing4.0-29B-A4B FP8 weights
• Xing4.0-29B-A4B GGUF weights
• China Telecom AI September 22 release announcement