Chinese large-language model (LLM) companies are adapting domestic computing power chips, a move that analysts said will accelerate the development of the artificial intelligence (AI) industry ecosystem. Chinese tech company Xiaomi released and open-sourced its MiMo-V2.5-Pro, while Chinese Graphic Processing Unit (GPU) maker MetaX announced that it had completed Day Zero ultra-fast adaptation of the model, allowing it to run smoothly on its Xiyun C-series GPU platform. “This adaptation marks a significant step forward in the collaborative development of domestic computing power and Chinese LLMs, jointly promoting the openness and prosperity of the global AI ecosystem,” MetaX noted. Chinese chip company Enflame said that it had also synchronously completed Day Zero ultra-fast adaptation for MiMo-V2.5-Pro, becoming one of the first domestic computing power vendors to achieve compatibility. Enflame noted that the adaptation further strengthens the foundational infrastructure for domestic AI computing power. Leveraging independently developed integrated software and hardware capabilities, it enables more efficient deployment of Chinese LLMs, provides strong support for the large-scale application of domestic AI, and jointly promotes the open integration and prosperous development of the global AI ecosystem.
The rapid ecosystem response highlights the accelerating momentum in China's domestic AI stack. On April 24, DeepSeek released its highly anticipated V4 series and achieved deep adaptation with Huawei's Ascend ecosystem, with Huawei announcing full support across its Ascend supernode product line for high-performance inference. This is not an isolated case. In recent months, many leading Chinese LLMs have taken proactive steps to deeply integrate with domestic computing hardware, forming a clear trend of close model-chip collaboration. For example, Step 3.5 Flash from StepFun was promptly adapted by multiple top chip vendors including Huawei Ascend, MetaX and Biren Tech. Similarly, models such as Zhipu GLM-5, Alibaba's Qwen series, and Baidu's ERNIE have also achieved broad compatibility with domestic platforms.
“This demonstrates that the training of China's top-tier LLMs can now be fully conducted on domestic computing platforms. Whether in mitigating risks related to computing power supply or achieving cost-effective deployment of LLMs, it holds significant strategic value,” Xiang Ligang, Director General of the Zhongguancun Modern Information Consumer Application Industry Technology Alliance, told the Global Times. Xiang noted that it fully showed that pairing domestic LLMs with domestic chips is no longer merely an industry aspiration or vision. It has become a practical reality with full readiness for commercial deployment and large-scale application, the Global Times reports.
