ERNIE 4.5
by Baidu·China·Released
Baidu's multimodal flagship — Chinese-language leader, integrated with the Baidu ecosystem.
About this model
ERNIE 4.5 (March 2025) is Baidu's flagship — a multimodal closed-weights model accessible via the Baidu Cloud Qianfan platform. ERNIE 4.5 introduced significant multimodal capability (text + vision + audio) over the previous ERNIE 4.0 generation, and is the primary AI engine behind Baidu's consumer products (Yiyan chat app, Baidu Search AI Overviews).
For Mainland China deployments where data residency and CAC content compliance matter, ERNIE 4.5 is often the path of least resistance — integrated with WeChat / Alipay / Baidu Maps, and supported by Baidu's enterprise sales organisation.
Strengths
- •Native Chinese-language specialist — best CN-language performance
- •Deep integration with Baidu Search and the broader Baidu ecosystem
- •CAC content-compliant by default for Mainland deployments
- •Multimodal: text, vision, audio in one model
Limitations
- •English-language quality trails US frontier models
- •Closed weights, Mainland-China data residency only
- •Limited international developer ecosystem outside Baidu Cloud
- •Less benchmark transparency than Western or even other Chinese labs
When to use it
- →Mainland China consumer apps integrating Baidu services
- →Chinese-language enterprise customer support
- →Search-augmented Chinese assistants
- →Public-sector deployments under CAC compliance requirements
Architecture & training
Baidu has confirmed ERNIE 4.5 uses a MoE architecture but has not disclosed parameter counts. Trained on a mix of Baidu's Kunlun chips and Nvidia hardware. The pretraining corpus heavily emphasises Chinese web content, search query logs, and structured knowledge from Baidu Baike. The ERNIE X1 reasoning variant (released around the same time) is built on the same backbone.
Benchmarks
| Benchmark | Score | Bar |
|---|---|---|
| CMMLU | 88.3 | |
| C-Eval | 89.8 |