With the launch of R1, DeepSeek not only created a shockwave in Silicon Valley but also intensified competition within the Middle Kingdom. Last February, Baidu, a major Chinese player in GenAI and the owner of the eponymous search engine, responded with the announcement of the open source release of its AI model Ernie next June, as well as the free availability of its chatbot Ernie Bot. They have doubled down since yesterday by publishing ERNIE 4.5, the latest version of its multimodal foundation model, and also ERNIE X1, a reasoning-focused model offering performances comparable to R1 but at half the price. Both are integrated into Ernie Bot, which is thus free a little earlier than expected.
According to Baidu, ERNIE 4.5 incorporates significant advances in understanding, generation, reasoning, and memory. It notably benefits from better management of hallucinations and optimization of logical reasoning. Its ability to simultaneously process text, images, sound, and video makes it a powerful tool for various applications, from dialogue to content creation.
The improvements to the model are attributed to several key technologies, including dynamic attention masking 'FlashMask', heterogeneous multimodal expert mixing, spatio-temporal representation compression, knowledge-centered training data construction, enhanced post-training self-feedback.
In its statement, the company claims that ERNIE 4.5 outperforms GPT-4.5 in several benchmarks, while being 100 times less expensive.
On the other hand, ERNIE X1, focused on multimodal reasoning and the use of advanced tools, excels in planning, analysis, and solving complex problems. It integrates specific features such as advanced search, image generation and interpretation, web page reading, and concept mapping via TreeMind.
A Bet on Accessibility and Competitiveness
By making these models free for the general public and offering competitive rates for businesses via its Qianfan cloud platform, Baidu aims to strengthen its influence in the AI ecosystem. ERNIE 4.5 is thus offered at 0.004 RMB (around 0.0005 euros) per thousand input tokens and 0.016 RMB (around 0.002 euros) in output, while ERNIE X1 has even lower prices: it is offered at 0.002 RMB (around 0.00025 euros) for the same number of input tokens and 0.008 RMB (around 0.001 euros) in output.
This aggressive pricing position is accompanied by a desire to democratize generative AI while gradually integrating these models into Baidu's products and services, including its search engine and the Wenxiaoyan application.
By offering a model with performances comparable to R1 but at half the price, Baidu clearly shows its ambition to dominate the sector not only in China but also against American players.