Baidu Intelligent Cloud Releases Three Lightweight Large Models: ERNIE Speed, Lite, and Tiny
-
Baidu Intelligent Cloud recently held a grand Qianfan product launch event, where it prominently introduced three lightweight large models: ERNIE Speed, ERNIE Lite, and ERNIE Tiny. Compared to large models with hundreds of billions of parameters, these lightweight models significantly reduce the number of parameters, making them more convenient for customers to fine-tune for specific application scenarios. This design not only helps customers achieve their desired usage outcomes more easily but also saves them substantial costs. These three lightweight large models each have distinct characteristics, with parameter sizes decreasing from ERNIE Speed to ERNIE Tiny. ERNIE Speed is suitable as a base model for fine-tuning in specific scenarios, balancing model effectiveness and inference performance. It can also be deployed on low-computing-power AI accelerator cards for inference, meeting the needs of low-cost and low-latency applications. Meanwhile, ERNIE Lite and ERNIE Tiny cater to a wider range of usage scenarios to varying degrees.
It is worth mentioning that Baidu Intelligent Cloud ModelBuilder has also launched two vertical-scenario large models: ERNIE Character and ERNIE Functions. These models are developed based on Baidu's extensive business experience and are tailored for role-playing applications (such as game NPCs and customer service dialogues) and tool-calling scenarios (e.g., using external tools or invoking business functions during conversations). Enterprises can directly apply these proprietary models to develop intelligent assistants without additional fine-tuning, significantly improving development efficiency and model applicability.
Particularly noteworthy is ERNIE Speed, Baidu's self-developed lightweight large language model. Its rapid response capability and minimal data fine-tuning requirements drastically reduce training time. In specific scenarios, ERNIE Speed's performance can even rival that of the ERNIE Bot 4.0 model. The model excels in natural language processing tasks such as text classification, named entity recognition, and semantic matching, especially in applications like intelligent customer service, search engines, and intelligent recommendations. Moreover, ERNIE Speed can achieve or even surpass the performance of trillion-parameter large models in complex tasks such as reading comprehension, closed-book question answering, and creative writing/continuation, fully demonstrating its strong application potential and value.