China Telecom's Xing4.0 Ships After Weeks of Leaked PRs

China Telecom published the weights for Xing4.0-29B-A4B on Hugging Face Wednesday, two days after the vLLM and SGLang serving frameworks both merged support for a model under that name, ending six weeks in which the checkpoint existed only as a trail of pull requests.

The model card describes a 29-billion-parameter mixture-of-experts model with 4 billion parameters active per token, a 256,000-token native context extensible to 512,000, and a mixture-of-experts routing scheme with 64 routed experts and one shared expert. China Telecom Artificial Intelligence Technology Co. calls it the successor to its TeleChat line and says it is the first model of this scale trained entirely on Huawei's Ascend 910C NPUs using the MindSpore framework, with multi-level co-optimization delivering roughly a 96 percent training-throughput gain over out-of-the-box performance. It ships under an Apache 2.0 license.

Six PRs, three names, six weeks

Before the weights existed publicly, the model was already visible through its code. OrcaRouter, a site that tracks unreleased models through open-source framework pull requests, documented six separate integration attempts across vLLM and SGLang between early August and Sept. 16, filed under three successive names: TeleChat4, then XingChen4, then Xing4_0. A vLLM pull request opened Aug. 27 carried the note "Model weights are not yet public on Hugging Face Hub"; the same caveat reappeared in a follow-up filed Sept. 16, hours before SGLang's own version added the first published parameter count, "29B-parameter MoE with ~4B activated parameters," plus a launch command naming a 262,144-token context and EAGLE speculative decoding.

The config files those PRs exposed describe an architecture that reuses DeepSeek's V2 and V3 backbone, including multi-head latent attention, and replaces the standard residual connection with Manifold-constrained Hyper-Connections, or mHC, a technique DeepSeek itself published in December and used in DeepSeek V4 that April. mHC expands a transformer's single residual stream into several parallel streams mixed by doubly stochastic matrices, keeping signals from exploding across hundreds of layers; DeepSeek reported roughly 6.7 percent training overhead for the technique at four residual streams. China Telecom's own model card does not mention DeepSeek or disclose the residual design, framing the release instead around its Ascend training claim.

Where it lands on benchmarks

China Telecom's own numbers put Xing4.0 ahead of Google's Gemma4-26B-A4B on every listed benchmark and mixed against Alibaba's Qwen3.6-35B-A3B, a larger model with a smaller fraction of parameters active, from the Qwen lineage that topped Hugging Face's download charts this month:

Benchmark	Xing4.0-29B-A4B	Gemma4-26B-A4B	Qwen3.6-35B-A3B
SWE-bench Verified	75.00	53.00	76.00
Terminal-Bench 2.1	57.50	30.00	51.50
Claw-Eval	76.55	71.49	74.54
DeepresearchBII	60.80	39.30	59.70

Those figures are vendor-supplied, run with the company's own harness settings and not yet reproduced outside China Telecom.

What's unverified

The Ascend training claim is China Telecom's own; OrcaRouter's tracking, drawn entirely from public pull requests and config files, says nothing about which hardware trained the model, only what the code says it computes. Whether the throughput and benchmark figures hold up under independent testing is now checkable for the first time since the weights are public, something the six weeks of PR-only signal could not offer. China Telecom has not said when, or whether, it will publish a technical report matching the detail of its TeleChat3-MoE training report from December.