According to Commercial Times, following the success of DeepSeek-R1 earlier this year, major Chinese tech companies have been actively advancing the development of open-source AI models, driving down the cost of using large models by 60% to 80%. Notably, on June 30, both Baidu and Huawei announced the release of their models as open source, as the report highlights.
Baidu goes open with Ernie 4.5
As noted by South China Morning Post, Baidu entered the Chinese open-source AI arena by making its flagship Ernie 4.5 models available for download on the AI platform Hugging Face. The report states that Baidu has open-sourced 10 variants from its Ernie 4.5 multimodal model family, ranging from lightweight models with 0.3 billion parameters to large-scale versions with up to 424 billion parameters.
Baidu’s move to open source marks a significant shift in strategy, as South China Morning Post indicates. The Beijing-based tech giant, one of the first in China to develop large language models (LLMs) after ChatGPT’s release in November 2022, has now reversed course—despite founder and CEO Robin Li Yanhong stating just a year ago that the Ernie series would surpass open-source alternatives.
Huawei launches open-source Pangu models, optimized for Ascend chips
In addition to Baidu, Huawei has open-sourced its Pangu AI models, including a 7-billion-parameter version and a 72-billion-parameter Pangu Pro MoE (Mixture-of-Experts) model, according to Tech in Asia.
The release also includes inference technology optimized for Huawei’s Ascend chips. Commercial Times highlights that this is the first time Huawei has open-sourced the core capabilities of its Pangu models.
China’s broader open-source momentum
As the open-source wave of large AI models gains momentum in China, companies such as MiniMax, Alibaba, and Moonshot AI have also been making notable strides, according to Commercial Times.
Commercial Times also mentions that in February, DeepSeek launched an “Open-Source Week,” during which it released one codebase per day over the course of a week—sharing its research progress with global developers.
By contrast, Commercial Times notes that U.S. AI giants such as OpenAI—whose GPT-1 through GPT-4 models remain closed-source—as well as Google and Anthropic, have all adopted proprietary approaches in their model development.
Source: TrendForce, Taiwan.
