metamorworks - stock.adobe.com
Baidu makes foundation model Ernie 4.5 open source
The China-based AI vendor has contributed to the open source community before. The release of the family of open models highlights the growth of the open source community in GenAI.
China-based AI vendor Baidu on Monday made its latest family of models, Ernie 4.5, open source.
The Ernie 4.5 family of multimodal models consists of 10 different variants. Baidu initially introduced the model family, in a non-open source version, in March. The models are MoE, or mixture-of-experts-based, with parameter sizes between 3B and 424B.
Baidu trained the family of models with the PaddlePaddle deep learning framework. Baidu initially developed PaddlePaddle, which allows for high-performance inference. The models are publicly accessible under the Apache 2.0 license. The models are optimized for general-purpose language understanding and generation. They are trained in both textual and visual modalities. Baidu said that what it calls the vision language models (VLMs) focus on visual language, understanding, thinking and non-thinking modes. The AI vendor also made the development toolkits for Ernie 4.5 open source.
Baidu and open source
The open source version of Ernie 4.5 is on brand for Baidu, as it has released other tools to the open source community before, like its PaddlePaddle framework. However, it also speaks to the continual growth of open source over the last year. Another Chinese vendor, DeepSeek, released its open source model, Deepseek-R1, earlier this year.
Both Baidu and DeepSeek compete against Chinese tech giant Alibaba and U.S.-based vendors such as OpenAI and Google. On June 25, Google introduced Gemini CLI, an open source AI agent. Elon Musk’s xAI is another AI vendor that has made its models open source. On Monday, Chinese tech giant Huawei also released its Pangu AI model to the open source community.
"Open source is going to dominate," said Mike Gualtieri, an analyst with Forrester. "I think that's good for everyone." He added that being known as a vendor with just models is not profitable. Chinese vendors have probably noticed this, which is why they are releasing their models to the open source community. Gualtieri said.
"They probably just see the writing on the wall anyway, and you can't just make money out of the model,” he said. “It's a good strategy to open source them.”
New trend and MoE
As for Baidu, the vendor is following a new trend in which, instead of just opening one large model, it's making a whole family of models open source.
Another aspect that makes Baidu stand out is its MoE architecture, said Bradley Shimmin, an analyst with Futurum Group.
Baidu said it designed a heterogeneous MoE structure that ensures that textual and visual modalities are represented for reinforcement during training.
"You see them doing some fascinating work with their mixture of expert architecture to share weights across the different modalities in the different areas of expertise so that you don't end up compromising as you do with some other mixture of experts models," Shimmin said.
"They're providing some value where we didn't know we could achieve before,” he continued. "They're trying to push forward what it means to be a frontier model."
A challenge
While open source has gained momentum in the past few years, one notable challenge for enterprises going the open source route is a lack of support.
While that challenge has not been resolved, some generative AI platforms such as Amazon Bedrock support both open and closed-source models.
Shimmin said that the appeal of open source is that it's community-driven. He added that some developers have already started to experiment with the new family of models.
"All that is the community of data scientists, ML engineers, researchers, and enterprise practitioners pledging to this project," Shimmin said. "The community wants to contribute, and that's what I think is so different about open source. It's not the company that has to provide the support. It's the community that does it."
Esther Shittu is an Informa TechTarget news writer and podcast host covering artificial intelligence software and systems.
