NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

China Telecom Built a Large AI Model Without Nvidia Chips — Then Gave Away a Smaller One

China Telecom's Xing4 paper details a 285B model trained on 8,192 Ascend chips, with an open 29B sibling — a full-stack experiment in training without CUDA.

2026-11-20 · 1186 words · NeuroAI
China Telecom Built a Large AI Model Without Nvidia Chips — Then Gave Away a Smaller One

Training a modern AI model has a default settings list, and near the top of it sits a stack of processors from one American company. For years that was less a choice than a fact of life. A team at China Telecom decided to find out what happens if you delete that line.

On November 16, researchers at the company's Xingchen AGI Lab — the unit also known as TeleAI — posted a technical paper with a blunt subtitle: advancing agentic models on domestic chips and frameworks. By November 18 it had climbed to the top of Hugging Face's Daily Papers list. The reason was not a new record on a chat leaderboard. It was the claim underneath: a frontier-scale training run with no Nvidia hardware and no CUDA anywhere in the pipeline.

Key takeaways

  • The team reports pre-training a 285 billion parameter mixture-of-experts model on 8,192 Huawei Ascend 910C processors using MindSpore HyperParallel. About 17 billion parameters activate per token.
  • That flagship is not released. What is public is Xing4.0-29B-A4B, a 29 billion parameter model with roughly 4 billion active parameters, published under an Apache 2.0 licence.
  • The paper reads like an engineering log: wider and shallower layers, reworked parallelism, a custom attention stack, and a speculative-decoding method called DSpark.
  • The model card says training throughput improved about 96% over an out-of-the-box baseline. That is a relative gain on one cluster, not a universal benchmark.
  • This is a research and infrastructure story. There is no consumer product and no evidence yet that the approach beats a GPU stack on price outside this specific setup.

A training log, not a launch splash

Most model releases arrive wrapped in a benchmark table. This one arrives wrapped in a confession. The authors spend much of the paper explaining that you cannot simply lift a CUDA-era training recipe, drop it onto a different cluster, and expect it to run. The smoother path is to design the model around the silicon you actually have.

They chose a wider, shallower arrangement of layers because, they report, it trains roughly 10% faster on Ascend clusters than the deep-and-narrow shapes that are common on GPU fleets. Attention combines several published ideas — multi-head latent attention from the DeepSeek lineage, sparse attention, and a technique the authors call IndexShare — and the report claims a 4.68-fold gain in prefill throughput over a baseline. For generation, a method named DSpark is said to deliver between 2.67 and 3.77 times the throughput.

None of those figures is a stand-alone product claim. They are the result of co-design: topology-aware sharding, distributed optimiser execution, coordinated memory management and CPU core isolation across an 8,192-card mesh. Removing pipeline parallelism alone is reported to have cut step time by a third — the kind of detail a press release skips and an engineer circles.

The post-training side has a name too. TeleGym is the team's platform for agentic training, built around asynchronous execution and partial rollouts so that tool calls of wildly different lengths can still be stitched into coherent training trajectories. An agent (智能体) that spends two seconds searching and forty seconds reasoning is awkward to train in a rigid batch; TeleGym is the answer to that awkwardness.

What the small model actually is

The 285 billion parameter flagship stays closed, which is the obvious caveat. The open member of the family is the one anyone can download: Xing4.0-29B-A4B. The naming describes the trade. Twenty-nine billion total parameters, about four billion lit up for each token. That is a sparse mixture-of-experts (混合专家) design aimed at long-horizon tasks, where a model must plan several steps, call tools, and keep a thread of reasoning intact.

The model card lists a native context window of 256,000 tokens, extendable to 512,000, and builds in BF16, FP8, W8A8 and GGUF formats. It is tuned for intent recognition, table understanding, contract review and knowledge questions — the unglamorous work that fills enterprise workflows. Fine-tuning runs through LLaMA-Factory or MindFormers; inference through vLLM, SGLang or KTransformers. A separate FlagOS adaptation layer means the same weights can be deployed across a range of domestic chip platforms.

There is a separate question of who this helps. A 29B-total, 4B-active profile is small enough to run on a single well-equipped machine after quantisation, and the licence is permissive. For a company that cannot send documents outside its own walls, a model it can inspect and host itself is worth more than a slightly higher score behind an API. The team appears to be betting that controllability and cost per completed task may matter more than raw parameter count.

Why the hardware story matters

The interesting claim is not that a Chinese telecom operator trained a model. It is that it trained one at this scale without the industry's default accelerator, and wrote down enough of the process for others to study. For readers outside China, the significance is less about geopolitics than about an engineering fact: the assumption that frontier training requires one specific vendor's stack has now been tested in public, with an unusually detailed log attached.

That does not mean the stack is finished. A model that runs is only the first step; a toolchain developers enjoy using is a different achievement. The authors note that long context does not guarantee reliable long tasks, and that tool calls touching payments or data deletion still need permissions and human review.

Honest limitations

This article draws on the technical report's abstract as reproduced by established AI trackers, the official model card on ModelScope, and English-language summaries of the paper. I have not read the full research report, so the architecture and throughput figures follow the team's own descriptions rather than an independent reading of the experiments.

The 96% training-throughput figure is measured against an out-of-the-box baseline on the team's own Ascend cluster. It is not a comparison against an equivalent GPU cluster, and it should not be read as a general speed advantage. The 4.68-fold and 2.67-to-3.77-fold numbers come from the same report and have not been replicated outside it. "Frontier-scale" is the authors' framing, not a verdict from an outside evaluator.

I have not verified the licence terms beyond the Apache 2.0 label, nor tested the weights on my own hardware. Translations of the lab's and model's names can differ across outlets.

What readers can do now

  • If you run models locally, download the 29B weights and test them on your own long-document task. Ignore the leaderboard; measure whether it finishes the work you actually care about.
  • If you are evaluating vendors, ask what hardware a model was trained and served on. The answer increasingly affects both cost and data-control options.
  • If you are simply watching the industry, treat this as a data point rather than a verdict. One well-documented training run is a beginning, not the end of an argument.

A model is a bundle of choices about what to build and what to leave out. What China Telecom has left out, conspicuously, is the chip everyone assumed was mandatory. Whether that becomes a pattern or a footnote depends on the next run, not this one.

Related coverage

Cite this article

NeuroAI (脑机网), "China Telecom Built a Large AI Model Without Nvidia Chips — Then Gave Away a Smaller One," 20 November 2026, https://neuroai.site/a/na-model-xing4-domestic-ascend-training

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Foundation Models” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories