Baidu releases ERNIE 4.5, a native multimodal foundation model

Background

The March 2025 release introduced ERNIE 4.5 as a native multimodal model and paired it with the reasoning model ERNIE X1.

What happened

On 16 March 2025 Baidu released ERNIE 4.5 together with the reasoning model ERNIE X1.^[1] ERNIE Bot became free to individual users the same day, ahead of the planned 1 April date.^[1]

Native multimodal means that a model is designed from the start to handle several kinds of input, rather than combining separate single-kind models afterwards. A reasoning model is one that works through a problem step by step before answering. API prices on platforms such as Qianfan are usually quoted per thousand or per million tokens, where a token is a short unit of text, often part of a word. A mixture-of-experts design activates only part of a model for each input, which can reduce the computation needed for each token. Open-source releases make a model's weights available to others, while the licence sets the terms of use.

Key details

Baidu called ERNIE 4.5 a native multimodal model that handles text, images, audio and video. It uses FlashMask dynamic attention masking, a technique for controlling which tokens attend to which during processing, and a heterogeneous multimodal mixture-of-experts (MoE) design.^[1]

Baidu claimed that ERNIE 4.5 outperformed GPT-4.5 on multiple benchmarks at about 1% of GPT-4.5's price.^[1]

Launch API prices on Qianfan started at RMB 0.004 per thousand input tokens and RMB 0.016 per thousand output tokens.^[1] VentureBeat converted these prices to $0.55 per million input tokens and $2.20 per million output tokens, against $75 and $150 for GPT-4.5.^[2]

Baidu said on X that it planned to open-source the ERNIE 4.5 family on 30 June 2025.^[2]

Impact

The release combined a native multimodal model, a reasoning model and a free consumer product in one announcement. VentureBeat set the API prices side by side with those of GPT-4.5.^[1][2]