ERNIE 5.0
ERNIE 5.0 is a 2.4-trillion-parameter mixture-of-experts foundation model. Its research architecture learns text, images, video and audio within a shared autoregressive system.
Architecture and access
The research report describes modality-independent expert routing and separate token-prediction schemes for language, vision and audio. Elastic training allows smaller configurations to reuse the larger model’s training. Qianfan lists an ernie-5.0 endpoint. Research demonstrations of media generation do not establish that every modality is exposed through that endpoint.ERNIE 5.0 unified multimodal research report
Practical use
Practical guidance. Choose the endpoint by the input and output types in its current API contract. A model that was trained on audio and video does not automatically provide an audio or video generation API. Verify the exact workflow before connecting it to a production content pipeline.ERNIE 5.0 unified multimodal research report
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | API |
| Model identifiers | ernie-5.0 |
| Hosted limits | 128K context; maximum output 65,536 tokensQianfan current model and endpoint list |
| Research activation | Below 3% of parameters per tokenERNIE 5.0 unified multimodal research report |
Related models
Hosted general-purpose model derived from ERNIE 5.0 through elastic model selection and task-specific post-training.
Launch date API · Provider product- ERNIE 5.0 Thinking Preview
Baidu · ERNIE 5
Hosted ERNIE 5.0 preview with reasoning enabled by default and an optional non-thinking mode.
Launch date not confirmedPreviewAPI