Xing4.0-29B-A4B

An agent model from a phone company, trained without Nvidia. Xing4.0-29B-A4B is an open-weight agentic MoE from China Telecom AI (the Xing series, formerly TeleChat). It does the agent loop: plan multi-step tasks, call tools, chew through long documents, hand back finished work. The distinguishing fact is the training hardware: the whole run happened on Huawei Ascend NPUs with the MindSpore framework, Ascend 910C clusters, no Nvidia cards anywhere in it. It is the first model of this size with a chips-to-framework all-domestic chain behind it.

What runs where. 29B total parameters, 4B active per token, 64 routed experts (4 active + 1 shared) with MLA attention, 256K context native, extensible to 512K. In bf16 the weights are 62.4 GB (server territory). The deployment story is quantization: the vendor says a 4-bit quantized build needs 15 GB of GPU memory, and the math agrees (29B x 4 bits is 14.5 GB), which puts it on one RTX 3090 or 4090 with room left for KV cache. A community GGUF (IQ4_NL, 20.1 GB file) is already on HuggingFace with 10,880 downloads. vLLM, SGLang, and KTransformers are supported for serving; LLaMA-Factory and MindFormers for fine-tuning.

Benchmarks, labeled. SWE-bench Verified 75.0 and Terminal-Bench 2.1 57.5 are vendor-run numbers from the model card (SWE-agent harness, 210K context window); Terminal-Bench 2.1 at 57.5 versus 30.0 for Gemma4-26B-A4B and 51.5 for Qwen3.6-35B-A3B is the standout row. SuperCLUE agent capability 93.52 puts it third, less than one point behind the top two Qwen models. Treat all of these as vendor-measured until third parties replicate.

In production, not just on a chart. China Telecom runs it in its group-level customer service platform for multi-step inquiry resolution, and in mid-screen interactive scenarios. The company says larger Xing models are coming.

The market signal. A state telecom shipping a competitive small agent model, open weights, Apache 2.0, is a statement about where compute sovereignty is going: model quality is now achievable without touching Nvidia silicon, and the release PR is aimed at developers running agents on their desktops.