ERNIE 4.5
Baidu's multimodal mixture-of-experts language model series.
ERNIE 4.5 is a family of large language models developed by Baidu, released in 2025. It includes text-only and vision-language variants built on a heterogeneous mixture of experts architecture. The models feature hundreds of billions of parameters and support open-weight distribution under the Apache License 2.0.
- Released by Baidu in 2025 under Apache License 2.0.
- Uses heterogeneous mixture of experts architecture.
- Activates 47 billion parameters per token.
- Supports both text and vision modalities.
- Available in PyTorch and PaddlePaddle formats.
What it is
ERNIE 4.5 is a suite of large language models created by Baidu. The series includes a text-focused variant with 300 billion total parameters and a vision-language model with 424 billion parameters. Both versions use a mixture of experts design, activating only 47 billion parameters per token. This approach allows for efficient processing while maintaining high capacity. The models are designed for general language tasks as well as complex visual reasoning.
How it came about
Baidu released the ERNIE 4.5 models in 2025. The company published technical reports detailing the architecture and training methods. These models represent an evolution in the ERNIE family, introducing heterogeneous expert structures to handle multiple data types simultaneously. The release included both PyTorch and PaddlePaddle weight formats to support different developer ecosystems.
Inside the model
The architecture uses a heterogeneous mixture of experts structure. This design separates text and vision experts to prevent one modality from interfering with the other. Training involves modality-isolated routing and specific loss functions to balance token representation. For inference, the system employs multi-expert parallel collaboration and quantization techniques. Post-training uses supervised fine-tuning, direct preference optimization, and unified preference optimization to align the models with user needs.
What it can do
The vision-language variant supports image understanding and cross-modal reasoning. It offers both thinking and non-thinking modes for complex tasks. The models can process web search references, assessing their relevance and authority before generating answers. They are optimized to provide professional responses by incorporating data from authoritative sources.
How to use it
The models are available with open weights under the Apache License 2.0. This license permits commercial use. Developers can download the weights from platforms like Hugging Face. The models are compatible with the transformers library and vLLM for inference. Baidu provides code snippets for loading the models in Python environments.
Questions people ask about ERNIE 4.5
What is the parameter count of ERNIE 4.5?
The text variant has 300 billion total parameters, while the vision-language model has 424 billion. Both models activate 47 billion parameters for each token processed.
Is ERNIE 4.5 open source?
Yes, the models are released with open weights under the Apache License 2.0. This allows for commercial use and modification according to the license terms.
What modalities does ERNIE 4.5 support?
The series includes a text-only model and a vision-language model. The latter can process both images and text, supporting tasks like image understanding and cross-modal reasoning.
Good reading and sources
There is no English Wikipedia article on ERNIE 4.5 yet, so the facts on this page come from the sources listed above: huggingface.co/baidu/ERNIE-4.5-300B-A47B-PT, huggingface.co/baidu/ERNIE-4.5-VL-424B-A47B-PT. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.