China Telecom's Xing4.0 Trains Entirely on Huawei Chips

China Telecom Artificial Intelligence Technology released Xing4.0-29B-A4B on Sept. 16, an open-weight mixture-of-experts model with 29 billion total parameters that activates only 4 billion of them on any given token.

The model uses 64 routed experts plus one shared expert, of which 4 activate per token, across 40 layers with a hidden size of 3,584, according to its Hugging Face model card. It is licensed Apache 2.0, and the weights are downloadable in Safetensors format with no gating. China Telecom also published FP8 and GGUF variants of the same model under the XingChen-AGI organization on Hugging Face, aimed at teams that want to run it on less memory than the full-precision release needs.

What it scores against Qwen and Gemma

On benchmarks China Telecom published with the release, Xing4.0-29B-A4B scores 75.00 on SWE-bench Verified, just behind Qwen3.6-35B-A3B's 76.00 despite having roughly 6 billion fewer total parameters, and well ahead of Gemma4-26B-A4B's 53.00 on the same test.

Benchmark	Xing4.0-29B-A4B	Qwen3.6-35B-A3B	Gemma4-26B-A4B
SWE-bench Verified	75.00	76.00	53.00
Terminal-Bench 2.1	57.50	51.50	n/a
Claw-Eval	76.55	n/a	n/a

On Terminal-Bench 2.1, a test of multi-step command-line tasks, Xing4.0 leads Qwen3.6 by 6 points, 57.50 to 51.50. On instruction-following, measured by IFBench, it scores 69.67, sitting between Qwen3.6's 65.50 and Gemma4's 72.67 on the related AA.LCR test at 61.00.

Why the training chip is the story

China Telecom said Xing4.0-29B-A4B is the first model of this scale trained entirely on Huawei's Ascend 910C NPU platform, using the MindSpore and MindFormers frameworks rather than Nvidia GPUs, according to MindStudio's benchmark report on the release. The model card describes roughly a 96 percent training-throughput improvement from mixture-of-experts-specific optimizations on that hardware, though it does not name the baseline configuration that comparison is measured against.

A related training report on Ascend infrastructure, posted to arXiv on Dec. 30, 2025 and led by Xinzhang Liu with 53 co-authors, describes the same TeleChat3-MoE family scaling from 105 billion to more than 1 trillion parameters entirely on Ascend clusters, citing near-linear scaling across thousands of devices as the reason the approach extends down to Xing4.0's smaller size.

What China Telecom left out

Neither the model card nor MindStudio's report gives a baseline number for the claimed 96 percent throughput gain, and China Telecom has not published a general-reasoning benchmark placing Xing4.0 against non-agentic tests the way it did for SWE-bench and Terminal-Bench. The model natively handles a 256,000-token context window, extensible to 512,000, and integrates with the agent frameworks OpenCode, Claude Code, OpenClaw and Hermes, according to the card and AlphaSignal's summary of the release, keeping the release squarely aimed at coding and terminal agents rather than general chat.

That framing puts Xing4.0 in the same lane as Qwen's single-GPU 27-billion-parameter release, competing on agentic coding benchmarks while avoiding the Nvidia-dependent supply chain that most rival labs still train on. The model supports deployment through vLLM, SGLang and KTransformers, according to the card, which covers the inference side even though training stayed inside Huawei's own stack.

What is missing from both documents is a like-for-like comparison against a Nvidia-trained model of similar size on the same benchmarks. Without that, the Ascend-only training claim is verifiable on its own terms, but the throughput advantage China Telecom cites has no independent baseline attached to it yet.