appConfigEntityId

bodyType

customContents

entityId

entityUuid

isVideo

articlePageQuery

article

articleSchemaContent

content

hooksArticlePageAdvertisementConfigArticle

hooksArticlePageFloatingButtonConfigArticle

appConfig

filter

json

liveArticle

Vincent Chow

State-owned China Telecom has developed the country’s first artificial intelligence models with the innovative Mixture-of-Experts (MoE) architecture that are trained entirely on advanced chips from Huawei Technologies.
This marked the first time a Chinese developer publicly validated the feasibility of using only Huawei chips to train AI models with MoE architecture, which has become widely adopted because of its ability to deliver high performance with fewer computational resources.
The TeleChat3 models – ranging from 105 billion to trillions of parameters – were trained on Huawei’s Ascend 910B chips and its open-source deep learning AI framework MindSpore, according to a technical paper published last month by China Telecom’s Institute of Artificial Intelligence (TeleAI).
The Huawei stack was able to meet the “severe demands” of training large-scale MoE models across many different sizes, TeleAI researchers said.
“These contributions collectively address critical bottlenecks in frontier-scale model training, establishing a mature full-stack solution tailored to domestic computational ecosystems,” they said.
The initiative by China Telecom, one of the world’s largest fixed-line and mobile network operators, underscored growing efforts by a number of US-blacklisted Chinese firms – including Huawei and iFlytek – to train AI models using domestically designed semiconductors.

Still, China Telecom’s self-reported performance scores for its TeleChat3 models showed that they lagged behind those of OpenAI’s GPT-OSS-120B, released in August, on several benchmarks.
That reflected the continued necessity to train cutting-edge AI models in China with graphics processing units (GPUs) from US suppliers Nvidia or Advanced Micro Devices.
AI models developed with the MoE architecture – which distributes tasks to multiple specialised submodels, or “experts” – can scale up capacity without significant increases in computational overhead.
MoE was popularised by DeepSeek’s V3 model, released in December 2024, and has since become the norm for leading-edge Chinese AI models.
MoE models, however, were considered more technically demanding than conventional models to train and run. They showed greater instability, especially at the fine-tuning stage, as many submodels must be managed.
Tsinghua University spin-off Zhipu AI – known internationally as Z.ai – last week said its new image-generation model was trained on Huawei chips, making it the first open-source model developed on an entirely domestic training stack to achieve industry-leading scores in image generation. Beijing-based Zhipu AI was blacklisted by Washington last January.
Last March, Ant Group researchers said they successfully trained a 300-billion parameter MoE model “without premium GPUs”, but did not specify whether they had exclusively used domestically designed chips. Ant Group is the fintech affiliate of Alibaba Group Holding, owner of the Post.
Nvidia, meanwhile, had pitched its suite of advanced GPUs and machine-learning frameworks as the world’s best tools for training large-scale MoE models.
Beijing, however, has made self-reliance across the entire AI stack a key priority for the country in the next five years, as US trade restrictions continued to block Chinese firms’ access to advanced US chips.
Although the US government recently gave the go-ahead for Nvidia to sell the H200, the firm’s second most powerful chip, to China, domestic tech companies have been told by Beijing that it would only approve H200 orders in exceptional circumstances, according to multiple reports.


Alibaba unveils ‘pragmatic’ AI road map to drive monetisation, infrastructure efficiency

Can the US and China verify an AI deal? Start-ups are starting to build the tools

Pacing problem: can AI fears overcome US-China race dynamics and force a slowdown?

China’s storage makers aim to climb value chain with advanced packaging bets

Semiconductors

China Telecom develops country’s first MoE models trained entirely on Huawei’s AI chips

A view of China Telecom’s stand at the annual MWC Barcelona trade show held in March last year. Photo: Shutterstock 

Huawei Technology’s Ascend 910B chip. Photo: Handout

Fighting back: Beijing builds its own ‘small yard, high fence’ to shut out US tech

Illustration: Henry Wong

National security adviser Jake Sullivan speaks during a press briefing at the White House in Washington, D.C., in this file photo dated October 1, 2024. Photo: Abaca Press/TNS

In this photo illustration, the Manus AI logo is seen displayed on a smartphone screen with the Meta logo in the background. Photo: SOPA Images/LightRocket via Getty Images

Beijing launched an investigation into back door risks of Nvidia chips in July. Photo: Shutterstock Images

Nvidia’s H200 AI chip. Photo: Handout

Chinese firms to incur higher costs for deploying advanced AI chips from US: analysts

The new import duties on advanced AI chips cover Nvidia’s H200 and Advanced Micro Devices’ MI325X graphics processing units. Photo: Shutterstock

The Bureau of Industry and Security, under the US Department of Commerce, this week gave Nvidia a conditional go-ahead to sell its H200 graphics processing units in China. Photo: Reuters

Why China may buy Nvidia’s H200 chips as Trump eases curbs – and replace them later

China’s domestic chip industry has made rapid advances in recent years, but it still lags the US in cutting-edge models. Photo: Shutterstock

Tech war: China semiconductor expert suggests vigilance in buying Nvidia’s H200 AI chips

Nvidia founder and CEO Jensen Huang speaks at the CES trade show in Las Vegas on January 6, 2026. Photo: AFP

Wei Shaojun, vice-president of the China Semiconductor Industry Association. Photo: Sina

Tech

Big Tech

Tech Trends

Tech War

Business

China Business

Companies

News

China Future Tech

The state-owned telecoms network operator trains its TeleChat3 models on Huawei’s Ascend 910B chips


The state-owned telecoms network operator trains its TeleChat3 models on Huawei’s Ascend 910B chips.


China Telecom develops country’s first MoE models trained entirely on Huawei’s AI chips

.css-1c6uqr6{color:inherit;font-weight:inherit;font-size:inherit;font-family:inherit;line-height:inherit;overflow-wrap:break-word;}The state-owned telecoms network operator trains its TeleChat3 models on Huawei’s Ascend 910B chips

The state-owned telecoms network operator trains its TeleChat3 models on Huawei’s Ascend 910B chips