ERNIE 4.5 VL 424B A47B

ERNIE 4.5 VL 424B A47B stands out as Baidu's flagship multimodal foundation model, built around a heterogeneous Mixture of Experts architecture designed from the ground up to process text and images together. Rather than treating visual and textual learning as separate challenges, the model uses modality-isolated routing and specialized loss functions—including router orthogonal loss and multimodal token-balanced loss—to keep visual and language representations from interfering with each other. With 128 total experts but only 16 active per token, the system delivers the representational power of a 424B parameter model while activating just 47B parameters during inference. The architecture incorporates grouped-query attention with 64 heads and 8 key-value heads across 54 layers, enabling sophisticated cross-modal reasoning, nuanced image understanding, and coherent long-context generation spanning up to 131,000 tokens.

The model's training pipeline reflects Baidu's investment in thorough post-training refinement. Beyond foundational multimodal pre-training, ERNIE 4.5 VL has undergone multiple rounds of alignment using Supervised Fine-Tuning, Direct Preference Optimization, and novel techniques like Unaligned Preference Optimization and Reinforcement Learning from Verbal Rewarding. This combination equips the model to operate in both reflective "thinking" mode and straightforward non-thinking inference, making it adaptable across different use cases and latency requirements. Built natively for vision-language tasks in English and Chinese, the model prioritizes efficient scaling and can operate under 4-bit or 8-bit quantization to reduce hardware demands. Released under an Apache 2.0 license with open weights, it invites developers and researchers to fine-tune, deploy, and build upon a foundation model that balances multimodal sophistication with practical deployment flexibility.