China Telecom Trains 285B LLM Without Nvidia
China Telecom's Xingchen AGI Lab published a technical report on Nov. 17, 2026, describing how it pretrained a 285B-parameter mixture-of-experts model on 8,192 Huawei Ascend 910C cards using MindSpore, without Nvidia GPUs or CUDA. Only the smaller Xing4.0-29B-A4B is released, under Apache 2.0.
China Telecom's Xingchen AGI Lab has detailed how it pretrained a 285-billion-parameter language model on 8,192 Huawei Ascend 910C chips, with no Nvidia GPUs or CUDA software in the loop. The technical report, "TeleChat to Xing4," was posted to arXiv on Nov. 17, 2026, and ranked as the No. 1 paper of the day on Hugging Face's Daily Papers on Nov. 18, after the AI Weekly newsletter flagged it in its daily briefing.

Key takeaways
- The Xing 4 report describes a 285B mixture-of-experts (MoE) model that activates about 17B parameters per token, supports contexts of up to 1 million tokens and was pretrained on 8,192 Ascend 910C cards.
- Only the smaller Xing4.0-29B-A4B is open-weight. It was released on Sept. 17, 2026, under Apache 2.0 and scores 75.0 on SWE-bench Verified, by the lab's own count.
- The lab reports up to 4.68 times higher prefill throughput from its attention design and 2.67 to 3.77 times faster decoding from speculative decoding, though some figures come from simulation.
Xing 4 285B: a large model trained without CUDA
The report notes that most open-source models are trained on GPUs and that scaling on domestic chips and frameworks "poses significant systems challenges." Xing 4 is the lab's answer: a series that extends China Telecom's earlier TeleChat models and runs on Huawei's MindSpore framework.
The 285B member is the report's main case study. It has 45 layers, 256 routed experts with eight selected per token, and one shared expert. Its first three layers are dense, and the rest use MoE blocks. The authors say production pretraining ran on 8,192 Ascend 910C cards drawn from a large-scale production cluster.
The report does not say the 285B weights will be released. It does say training of a trillion-parameter model is ongoing.
How the lab tuned for Ascend 910C chips
The authors say the Ascend 910C is strong at matrix math but relatively constrained in vector compute and memory bandwidth, so they shaped the model around that. A "wider-and-shallower" 45-layer design replaced a 60-layer baseline and improved training throughput by about 10%, they report.
Their training stack, MindSpore HyperParallel, shards model weights inside each Huawei SuperPoD so that pipeline parallelism, a common technique that leaves chips idle while they wait, can be dropped. Removing it raised throughput by 16.1%, according to the report.
On inference, combining multi-head latent attention (MLA) with DeepSeek Sparse Attention and a cross-layer index-sharing step delivered up to 4.68 times the prefill throughput of a plain MLA baseline at a 256K context. The authors caution that this figure comes from Huawei's Zhanlu simulator and may be more optimistic than real deployments. A speculative decoder called DSpark raised decoding throughput 2.67 to 3.77 times across six math and coding benchmarks.
Xing4.0-29B-A4B benchmarks
The open model has 29B total parameters, 4B active per token and a native 256K context, extensible to 512K, according to its GitHub repository. The lab says it is the first model of its size trained entirely on Ascend with MindSpore.
| Benchmark | Xing4.0-29B-A4B | Qwen3.6-35B-A3B | Gemma4-26B-A4B |
|---|---|---|---|
| SWE-bench Verified | 75.0 | 76.0 | 53.0 |
| Terminal-Bench 2.1 | 57.5 | 51.5 | 30.0 |
| SWE-bench Multilingual | 66.0 | 67.2 | 51.0 |
These scores have not been independently reproduced. The weights run on standard tools including vLLM, SGLang and KTransformers, and the lab credits DeepSeek's architecture work for improving training stability.
Context: China's push for a non-Nvidia stack
Chinese labs have leaned heavily on open-weight releases, including Alibaba's Qwen3.8 family, but most of those models were trained on Nvidia hardware. A full pretraining run of this size on Huawei silicon is a data point for whether China can build large models as U.S. export rules limit access to American AI chips.
What's next
The lab says the Xing 4 series will expand toward trillion-parameter scale on Ascend and MindSpore. Outside researchers will likely test whether the 29B model's agent scores hold up, and whether the report's efficiency gains carry over beyond China Telecom's own cluster.
FAQ
Can developers download the 285B Xing 4 model? No. Only Xing4.0-29B-A4B and its FP8 and GGUF versions are public, under the Apache 2.0 license.
Does Xing4.0-29B-A4B need Huawei hardware to run? No. The model card lists Transformers, vLLM, SGLang and KTransformers support, and GGUF builds run on CPU or CUDA.
Are the speed gains independently verified? No. They come from the lab's report, and the attention results are simulation-based.
References
- TeleChat to Xing4: Advancing Frontier Agentic Models on Domestic Chips and Frameworks — Xingchen AGI Lab, China Telecom, via Hugging Face Papers / arXiv (Nov. 17, 2026)
- GitHub - XingChen-AGI/Xing4.0-29B-A4B — China Telecom Artificial Intelligence Technology Co.
- AI News for November 18, 2026 — Daily Edition — AI Weekly (Nov. 18, 2026)
Editor responsible for AI Newsway. Writes about AI, developer tools, cloud and productivity technology.
Comments 0
Comments are currently disabled.