iFlytek Showcases AI Model Trained Entirely on Huawei Chips Amid US Tech Restrictions

Chinese voice-recognition company iFlytek announced that its Xinghuo X1 reasoning model, trained entirely using Huawei Technologies’ AI chips, has achieved performance on par with OpenAI o1 and DeepSeek R1.

Credit: iFLYTEK

The company said the model is “self-sufficient” and “controllable,” and was developed without reliance on Nvidia chips, which are now subject to tighter US export restrictions.

iFlytek and Huawei began co-developing the Xinghuo X1 model in June last year, aiming to overcome limitations in domestic chip interconnect bandwidth.

At the end of last year, Huawei’s Ascend 910B AI chip operated at only 20% of Nvidia’s efficiency for training reasoning models. That figure has since increased to nearly 80%, according to iFlytek founder and Chairman Liu Qingfeng.

Liu said during a Tuesday earnings call that the company’s reliance on Chinese computing platforms gives it a strategic edge as global organisations seek to reduce geopolitical risk.

He added that US restrictions on China’s access to computing chips will likely become more severe, making self-reliant AI models increasingly valuable for state-owned enterprises and critical industries.

Liu also noted that if China is fully cut off from Nvidia chips, iFlytek’s domestic infrastructure would serve as a vital safeguard.

The company plans to promote its large language models (LLMs) internationally, citing growing global interest in alternatives amid ongoing trade tensions.

In July 2024, iFlytek established its international headquarters in Hong Kong to offer LLM and voice recognition tools to local clients and expand its global presence in intelligent education hardware and services.

On Tuesday, the Shenzhen-listed firm reported 2024 revenue of 23.34 billion yuan (USD 3.1 billion), an 18.79% increase year on year. However, net profit declined by 14.78%, and shares closed 0.44% lower.

---

• iFlytek’s Xinghuo X1 model matches OpenAI o1 and DeepSeek R1 in performance

• Model trained entirely with Huawei’s Ascend 910B AI chips

• Efficiency of Huawei chips improved from 20% to nearly 80% of Nvidia’s

Source: SCMP