Baidu launches new AI models, touts superiority to DeepSeek, OpenAI on benchmarks

Tech giant says reasoning model Ernie X1 ‘delivers performance on par with DeepSeek R1 at only half the price’

Tech giant Baidu has launched new large language models (LLMs) and touted their superiority to DeepMind’s latest open-source model R1 and models from OpenAI on benchmarks.

The search giant is lagging behind in the open-source LLM race in China, but it highlighted the competitive pricing of its models and the performance edge on reasoning tasks.

The company unveiled its new LLM family, named Ernie X1, which includes a coding model, a multimodal model and a reasoning model.

The coding model can perform tasks such as generating code, translating programming languages and debugging code in multiple programming languages, while the multimodal model handles tasks like generating different creative text formats, image generation and text-driven video generation.

The company said that the reasoning model, specifically, “delivers performance on par with DeepSeek R1 at only half the price”.

Pricing for business access to Ernie X1’s application programming interface (API) is set at 2 yuan (US$0.28) per million token inputs and 8 yuan per million token outputs, the company said.

For comparison, OpenAI’s models are priced at US$0.07 per 1,000 tokens for input and US$0.08 per 1,000 tokens for output.

At the same time, DeepMind’s latest open-source model R1 has seen a surge in popularity among business and consumer users since its launch last month.

It also saw competitors such as Alibaba, Tencent and Bytedance unveiling their own LLMs and open-sourcing them to varying degrees.

Baidu was the first Chinese Big Tech firm to launch an LLM in China in March 2023 following the frenzy created by the release of OpenAI’s ChatGPT. But its early-mover advantage was challenged by China’s other AI players in the last two years.

The search giant’s latest move to increase its relevance in China’s AI market came as DeepMind has fanned an open-source wave, with several models from the likes of Union Tech and BOGO Labs also entering the fray.

The company did not provide benchmark results for its new reasoning model Ernie X1, but said that it "delivers performance on par with DeepSeek R1 at only half the price".

It added that the model also outperformed OpenAI’s models in terms of reasoning ability.

"In internal testing, Ernie X1 demonstrated superior performance in reasoning tasks compared to OpenAI models of the same generation, showcasing its edge in handling complex tasks and delivering nuanced outputs," Baidu said.

The company said that Ernie X1's performance was benchmarked against the GPT-4 series, highlighting its competitive edge in the rapidly evolving LLM landscape.

"In the realm of reasoning, Ernie X1 achieved performance on par with DeepMind's R1 model on the Stanford Reasoning Chain benchmark at half the cost, signifying its efficiency in performing sophisticated tasks such as summarization, translation and question answering."

The company said that the model's performance was benchmarked against the GPT-4 series, highlighting its competitive edge in the rapidly evolving LLM landscape.

"In the realm of reasoning, Ernie X1 achieved performance on par with R1 on the Stanford Reasoning Chain benchmark at half the cost, signifying its efficiency in performing sophisticated tasks such as summarization, translation and question answering."

The company said that the model's performance was benchmarked against the GPT-4 series, highlighting its competitive edge in the rapidly evolving LLM landscape.

"In the realm of reasoning, Ernie X1 achieved performance on par with R1 on the Stanford Reasoning Chain benchmark at half the cost, signifying its efficiency in performing sophisticated tasks such as summarization, translation and question answering."

The company said that the model's performance was benchmarked against the GPT-4 series, highlighting its competitive edge in the rapidly evolving LLM landscape.

"In the realm of reasoning, Ernie X1 achieved performance on par with R1 on the Stanford Reasoning Chain benchmark at half the cost, signifying its efficiency in performing sophisticated tasks such as summarization, translation and question answering."

The company said that the model's performance was benchmarked against the GPT-4 series, highlighting its competitive edge in the rapidly evolving LLM landscape.

"In the realm of reasoning, Ernie X1 achieved performance on par with R1 on the Stanford Reasoning Chain benchmark at half the cost, signifying its efficiency in performing sophisticated tasks such as summarization, translation and question answering."

The company said that the model's performance was benchmarked against the GPT-4 series, highlighting its competitive edge in the rapidly evolving LLM landscape.

"In the realm of reasoning, Ernie X1 achieved performance on par with R1 on the Stanford Reasoning Chain benchmark at half the cost, signifying its efficiency in performing sophisticated tasks such as summarization, translation and question answering."

The company said that the model's performance was