ERNIE 4.5 open-source family
Baidu open-sourced ten ERNIE 4.5 models under Apache 2.0, from a 0.3B dense model to a 424B-total (47B active) multimodal MoE.
A reversal for Baidu's closed Ernie line. Heterogeneous MoE with modality-isolated routing, router orthogonal loss and multimodal token-balanced loss; trained on PaddlePaddle. Includes 300B-A47B and 21B-A3B text models and 424B and 28B VL models.
- Date
- Monday, 30 June 2025
- Lab
- Baidu
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Variants released | 10 | company |
| Largest model (total / active) | 424B / 47B VL | company |
HF weights uploaded 2025-06-28 UTC; announcement 2025-06-30. Benchmark claims (e.g. 300B base beating DeepSeek-V3 base) are Baidu's.
Sources
This record was checked against its sources on 6 October 2026. How we check