Baidu open-sources the ERNIE 4.5 model family under Apache 2.0

Background

Before mid-2025, Baidu had kept its flagship ERNIE models closed, meaning they were offered through Baidu's own products and platforms rather than published for anyone to download. Baidu reversed that approach on 30 June 2025. Baidu had announced the open-source date when ERNIE 4.5 launched in March 2025, so the June release was a planned step rather than a surprise.^[5]

An open-weight model publishes its trained parameters, often called weights, so that developers can download the model, run it on their own hardware and adapt it. The license attached to the weights decides what users may do with them. Apache License 2.0 is a permissive license that allows commercial use, modification and redistribution, subject to conditions such as keeping the copyright notice.

What happened

On 30 June 2025 Baidu open-sourced ten ERNIE 4.5 models under the Apache License 2.0.^[1][2][3] The family spans a range of sizes. It starts with a 0.3B dense model, which uses all of its parameters for every token, and continues to mixture-of-experts (MoE) models of 21B (3B active), 28B (3B active), 300B (47B active) and 424B (47B active) parameters. The family also includes vision-language versions that can read images as well as text.^[3]

In a mixture-of-experts model, a router sends each token to only some of the network's expert sub-networks. The active parameter count is the part that runs for a given token, so a model such as the 300B-A47B version stores 300 billion parameters but uses about 47 billion of them for each token. This design aims to keep the knowledge of a large model while limiting the computation needed per response.^[3]

Key details

Baidu said that ERNIE-4.5-300B-A47B-Base beat DeepSeek-V3-671B-A37B-Base on 22 of 28 benchmarks. The Decoder noted that the models did not lead overall against the top models from OpenAI, Google or DeepSeek.^[3][2] The models were trained with PaddlePaddle, Baidu's deep-learning framework, and Baidu also released the ERNIEKit and FastDeploy toolkits for training and deployment.^[1][2]

Baidu added the open ERNIE 4.5 models to its Qianfan MaaS platform, a model-as-a-service offering for developers, and highlighted their instruction following, world knowledge and visual understanding.^[4]

Impact

The release made the ERNIE 4.5 weights available to developers outside Baidu's own products, under a license that allows commercial use. It also gave researchers a set of models at several sizes, including MoE versions with small active parameter counts, to study and compare.^[1][3]