China Telecom has trained its TeleChat3 AI models entirely on Huawei’s Ascend 910B chips, marking the first public validation of using only Huawei chips for MoE architecture models in China.

The models, which range from 105 billion to trillions of parameters, were developed using Huawei’s open-source framework MindSpore.

The company’s technical paper indicates that Huawei’s chips met the demands of training large-scale MoE models across various sizes.

However, China Telecom’s models reportedly lag behind OpenAI’s GPT-OSS-120B on several benchmarks.

🔗 Source: South China Morning Post