Baidu AI Cloud has boosted its Qianfan platform with five new Ernie models and an enhanced development toolkit, making advanced AI technologies more accessible and cost-effective for enterprises
The new Ernie models feature fewer parameters than large language models and deliver powerful performance across a range of applications with lower inference costs. This gives enterprises a more agile and efficient solution for customising AI to their specific needs.
Ernie Speed understands context, allowing it to generate more coherent and accurate content. It can reason well in a context window size of up to 128,000 tokens. For specific domains, the performance of Ernie Speed can reach a level on par with that of a flagship Ernie foundation model.
Ernie Lite is an upgraded version of Ernie Bot Turbo. Suitable for inference on low-computing power AI accelerator cards, it offers 20 percent higher performance in tasks such as sentiment analysis, multi-task learning and natural reasoning, and a 53 percent reduction in inference costs.
Ernie Tiny, the smallest model among the three, is ideal for customers seeking affordability and low latency, especially in applications such as retrieval, recommendation and intent recognition. In tests focused on conversational recommendations, Ernie Tiny excels in generating search engine keyword suggestions. It achieves a 3.5 percent boost in dialogue interactions compared to ERNIE 3.5, while reducing costs by 32 percent.
Ernie Character is created for scenarios based on roleplaying, such as allowing users to create non-playable characters in gameplay or customer service.
Ernie Functions is designed for customers requiring the integration of external tools or business functions. Enterprises can now skip additional fine-tuning and directly apply these proprietary models to develop their own smart assistants.
A toolkit upgrade for Qianfan AppBuilder includes 55 new tools aimed at simplifying the development of AI-native applications. It offers everything from core AI capabilities and Baidu’s proprietary technologies to advanced features such as retrieval augmented generation and generative business intelligence.
