Baidu launches new generation of Ernie AI

The AI marketspace is getting mighty crowded, and Chinese company Baidu is the latest player to launch its newest model into the world. At its Baidu World conference this week, it unveiled Ernie 5.0.

Baidu CTO and head of AI Group Haifeng Wang said (via translated subtitles supplied by the conference) that Ernie 5.0’s technical route was to “adopt a unified auto-regression architecture for native full multimodal modelling.” He said that this meant that “from the beginning of training, speech had been integrated, and images had been integrated, video, audio, and other multimodal data.”

While its predecessor, Ernie-4.5-VL-28B-A3B-Thinking, is supplied under an Apache license and is expected to provide an alternative to the likes of OpenAI, Ernie 5.0 is proprietary, built on the company’s PaddlePaddle deep learning framework.

In a written statement, Baidu said that Ernie 5.0 Preview was already in joint second place on the LMArena Text leaderboard, a well-established industry benchmarking organization (it has since fallen to eighth spot).

This claim was met with some skepticism by Thomas Randall, a research director at Info-Tech Research Group. “In my opinion, Ernie’s update is not a global game-changer. Claims of ‘outperforming’ or ‘matching’ other AI models must still be validated broadly in independent benchmarks and multiple languages/modalities. Having multimodal is good, but actual performance matters and we just don’t know yet whether Ernie substantially differs from other models.”

This emphasis on a multimodal approach is supported by Brandon Hoff, research director at  IDC. He said that Baidu’s focus on visual and STEM reasoning and visual grounding required a specialist approach. “These are all different workloads and targeted at different use cases, and they are getting smarter.”

He contrasted this to the OpenAI approach: “OpenAI is LLM chatbots that can address a wide variety of prompts.”

Hoff went on to say that this is an example of how quickly the AI world is moving. “I would categorize this announcement as one of many that we will see as AI is applied to new use cases and I would say that they are, in a way, expected or on the development curve,” he said. “Note that China has about half of the AI developers in the world, so we should expect lots of innovation coming out of China.”

However, AI researcher Ahmed Harhara, founder at Houston Home Tools, said that Baidu’s approach seemed heavily optimized for Chinese language context and local data. “It’s similar to how Google fine-tunes Bard [Gemini] for English-heavy datasets. The real difference isn’t just the model architecture; it’s the training environment and regulatory constraints each company operates under.”

He added that the importance of geographical fine-tuning in AI was underestimated. “It matters more than people think. It shapes how models interpret nuance, bias, and context. Baidu’s edge will likely be its integration with China’s digital ecosystem and search infrastructure, not just raw model performance.”

At the conference, Baidu EVP Dou Shen also announced two new processors aimed at powering the company’s advances in AI. The Kunlunxin M100, optimized for large scale AI inference, will be released at the beginning of 2026, with the M300, optimized for the training and inference of ultra-large-scale multimodal large models, following in early 2027.

In addition to these two new processors, Baidu also unveiled the Tianchi 256 and Tianchi 512 supernodes at the conference. They are expected to officially launch in 2026, with a single Tianchi 512 supernode, consisting of 512 Kunlunxin P800 chips, capable of training models with up to one trillion parameters, the company said.

Hoff noted that, in designing chips for their own workloads, the company was following in the footsteps of the US hyperscalers.  “Since they don’t have access to US AI platforms, China companies will focus on locally developed technology and are expected to do well,” he said.

The restrictions that the US has placed on exports to China is having an effect, he added, as it meant that China was now building on Chinese, not US, hardware platforms. “This is a threat to US leadership. Huawei is expected to be successful with their solutions, and they have tight relationships with many countries around the world, I would expect that Huawei and Baidu to both be strong competitors to the US in countries outside the US.