ERNIE 4.5

ERNIE 4.5, released by Baidu on March 16, 2025, is a multimodal AI model capable of processing text, images, audio, and video simultaneously. It represents a significant step in Baidu's AI development, offering both proprietary API access and open-source model weights for the broader community.

ERNIE 4.5 scores 79.6 in text understanding and general knowledge benchmarks, slightly outperforming GPT-4o (79.14). The ERNIE-4.5-VL-28B-A3B-Thinking variant, released in November 2025, achieves benchmark wins over GPT-5 and Gemini 2.5 in visual reasoning and document analysis tasks, using a compact 28B total / 3B active parameter design. A detailed technical report is publicly available.

Baidu subsequently released ERNIE 5.0 in January 2026 with approximately 2.4 trillion parameters (less than 3% active per query). ERNIE 5.0 scored 1,460 on LMArena, ranking 8th globally and 1st among Chinese models, on par with GPT-5.1 and ahead of Gemini 2.5 Pro and Claude Sonnet 4.5. It ranked 2nd worldwide in mathematics, trailing only GPT-5.2.