Baidu to open-source its Ernie large language model series – News

Baidu Inc. intends to open-source its Ernie series of large language models later this year.

The company, the operator of China’s most popular search engine, announced the plan today. Reuters reported that the initiative centers on an upcoming family of Ernie models dubbed the Ernie 4.5 series. Baidu will start rolling out the LLM series to users in the coming months and plans to open-source its code on June 30.

The company uses Ernie to power a ChatGPT-like chatbot service called Ernie Bot. It can generate text, solve math problems and perform other tasks. The service reportedly had 430 million users as of November, up from more than 200 million in April. 

Baidu released the first version of Ernie Bot in March 2023. At the time, the service was based on an LLM dubbed Ernie 3.0 that the company had started developing in 2019. The model featured 10 billion parameters and was trained on a 4-terabyte dataset.

The current version of the LLM, Ernie 4.0, is described by Baidu as a “full upgrade with drastically improved performance.” It can generate not only text but also images and videos. Additionally, the model has improved reasoning features that make it better at tasks such as solving geometry problems.

Ernie 4.5, the upcoming version of the LLM that Baidu plans to open-source, will reportedly be followed by an even more capable iteration called Ernie 5. A source told Reuters that the company plans to release the latter model in the second half of 2025. It’s unclear what features Ernie 5 is set to provide and whether it will also become available under an open-source license.

The shift in Baidu’s artificial intelligence strategy comes a few weeks after the launch of DeepSeek-R1 brought new attention to open-source LLMs. The model, which was created by Chinese AI lab DeepSeek, can outperform OpenAI’s o1 reasoning LLM across several tasks. Its release lead to a brief selloff in Nvidia Corp.’s shares. 

Baidu is not the only LLM provider that is sharpening its focus on open-source models in the wake of DeepSeek R1’s release. In a Jan. 31 Reddit thread, OpenAI Chief Executive Officer Sam Altman wrote that the company will ”need to figure out a different open-source strategy.” But he added that “not everyone at OpenAI shares this view, and it’s also not our current highest priority.”

OpenAI has already open-sourced several AI models, including the first two algorithms in its flagship GPT series of LLMs. The company’s GitHub page also hosts the code of multiple internally-developed programming tools. Those tools ease tasks such as creating simulations for training robots’ AI software.

Photo: N509FZ/Wikimedia