Baidu unveils the omni-modal ERNIE 5.0 at Baidu World 2025

Background

ERNIE 5.0 was presented as the next generation of Baidu's ERNIE large language models, following the ERNIE 4.5 family and the reasoning models X1 and X1.1. A foundation model is a large model trained on broad data so that it can be adapted to many tasks, and each new generation is usually judged on how much it can do and how many kinds of input it can handle.

The term omni-modal describes a model that works with several types of data, such as text, images, audio and video, within one system. A model described as natively omni-modal was trained to handle these types from the start, rather than having separate models for each type joined together afterwards.

What happened

Baidu unveiled ERNIE 5.0 at Baidu World 2025 on 13 November 2025.^[1] The company described the model as natively omni-modal across text, images, audio and video.^[1] A preview opened to the public on ERNIE Bot, and to enterprises on Qianfan, Baidu's model platform for developers and businesses.^[1]

Reports put the model at 2.4 trillion parameters, about twice the size of Qwen3-Max and Kimi K2.^[2] At the same event Baidu previewed the Kunlun M100 and M300 chips.^[2]

Key details

Robin Li said that "the speed of technological iteration is the only competitive moat."^[1] A moat is a lasting advantage that makes a business hard to copy, so the remark tied the pace of model development to Baidu's competitive position.

Baidu also launched GenFlow 3.0, which it said had more than 20 million users, and MeDo, an international no-code app builder.^[1]

ERNIE 5.0 exited preview in January 2026 and ranked eighth on the LMArena text leaderboard, making it the only Chinese model in the top 10.^[3] Baidu published the ERNIE 5.0 Technical Report on arXiv on 4 February 2026.^[4]

Impact

ERNIE 5.0 extended Baidu's model line from text-focused releases into a single system that accepts and understands several media types. Its placement made it the highest-ranked Chinese model in the top 10 of that text leaderboard.^[3] The technical report gave researchers a published description of the model.^[4]