Baidu’s new ERNIE model is making waves in the AI landscape, outperforming OpenAI’s GPT and Google’s Gemini in key benchmark tests. Specifically, the multimodal AI demonstrates significant advantages in processing and interpreting enterprise data. What’s more, this high performance is achieved with a surprisingly lightweight design, activating only three billion parameters during operation, a move aimed at reducing the inference costs that often limit the scalability of AI solutions.
ERNIE’s architecture integrates visual grounding with tool utilization, unlocking new potential for automation in various industries. The model particularly shines in its ability to interpret dense, non-textual data, showcasing its strength in technical domains. Benchmark results reveal that ERNIE-4.5-VL-28B-A3B-Thinking surpasses GPT-5-High and Gemini 2.5 Pro in challenging tasks like MathVista, ChartQA, and VLMs Are Blind evaluations. Furthermore, ERNIE is engineered to seamlessly manage external tools and autonomously enhance images to decipher small text, showcasing its versatility.
Available under the Apache 2.0 license, ERNIE is positioned to drive innovation in business intelligence by enabling the extraction of valuable insights from video archives and other complex data sources. Its commercial-friendly licensing model makes it an attractive option for businesses looking to leverage the power of advanced AI.