China Telecom Ships a 29B Model Trained Only on Huawei Chips
A.I. / news
China Telecom Ships a 29B Model Trained Only on Huawei Chips
Xing4.0-29B-A4B activates just 4 billion of its 29 billion parameters per token, fits on one consumer GPU, and was trained without a single Nvidia chip, the company said.
China Telecom AI released Xing4.0-29B-A4B on Sept. 22, 2026, a mixture-of-experts model, a design that activates only a fraction of a model's parameters for each token instead of all of them, that the company said was trained entirely on Huawei's Ascend NPU chips rather than Nvidia hardware.
The model card lists 29 billion total parameters with 4 billion active per token, a native context window of 256,000 tokens that extends to 512,000, and an Apache 2.0 license that permits commercial use. China Telecom AI said the training run used Ascend NPU clusters and Huawei's MindSpore and MindFormers software stack, with no Nvidia hardware in the pipeline, according to the model card.
With 4-bit quantization, the model needs 15 gigabytes of GPU memory, low enough to run on one consumer graphics card, the model card says. The company built the model on what it calls an mHC plus MLA plus MTP architecture and said co-optimizing the model against the Ascend hardware lifted training throughput by about 96 percent over its own baseline. It did not name that baseline configuration.
What the vendor's own benchmarks claim
China Telecom AI's own testing puts Xing4.0-29B-A4B at 75.0 on SWE-bench Verified, a benchmark built from real GitHub issues, and 90.0 on AIME2026, a competition-math benchmark. Neither figure has been reproduced by an independent evaluator such as Artificial Analysis as of Sept. 27, 2026.
| Benchmark | Score | What it measures |
|---|---|---|
| SWE-bench Verified | 75.0 | Real GitHub issue fixes |
| AIME2026 | 90.0 | Competition math problems |
| Claw-Eval | 76.55 | Agentic tool-use tasks |
| DeepresearchBII | 60.8 | Multi-step research tasks |
| Terminal-Bench 2.1 | 57.5 | Command-line agent tasks |
Where the company says it already runs the model
China Telecom AI said it has folded Xing4.0-29B-A4B into its own customer service platform for call routing and technical troubleshooting, and that related deployments now cover more than 500 enterprise workflows across 20 industry sectors, according to Telecoms Tech News. The company did not provide a way to check the workflow count independently.
"We believe the future of AI lies not in ever-larger models, but in making powerful intelligence accessible to everyone, affordable, and deployable anywhere," a China Telecom AI spokesperson said.
What the release leaves out
Xing4.0 follows China Telecom's TeleChat line of models; the company points to two earlier technical reports, one on TeleChat3-MoE from December 2025 and one covering TeleChat2, TeleChat2.5 and T1 from July 2025, rather than publishing a new paper alongside this release. It also did not name the specific generation of Ascend chip used in training, referring only to the Ascend NPU family and the MindSpore stack. For a company selling cloud and AI services inside China, where Alibaba's Qwen3.8-27B already runs on a single GPU, Xing4.0-29B-A4B is a bet that the chip story matters as much as the benchmark scores.
Anyone testing the hardware claims against Qwen3.8-27B's own published requirements will find both models converging on roughly the same consumer-GPU target, using different chip families to get there.
Sources
More in A.I.
- 01Reflection Announces Beam, a 501B Open-Weight Model, but Has Not Named a LicenceThe weights are promised for later in October. Reflection's own post gives scores and training scale, and says nothing on licence terms or API pricing.
- 02Meta Disputes Inc. Columnist's Claim That Muse Read His Messages While Mac Researcher Calls It a BackdoorTwo separate Muse problems are being reported as one. A researcher's local-access flaw is documented; the claim that the agent read private messages is contested by Meta.
- 03Anthropic Staff Reported a Claude Chat to Police, and a Florida Woman Faces a Felony ChargeThe company's privacy policy allows disclosure to prevent serious harm. Its transparency report counted zero emergency requests from police, and does not count referrals it makes itself.
- 04Vals AI's 90 Claude Opus 5.5 Agents Name Two Magnetic Semiconductor Candidates, Neither Yet MeasuredOne candidate was first synthesised in 1999 and has a measured ordering temperature of 376 K. The other may not survive the furnace.