ERNIE 4.5 VL 28B A3B

ERNIE 4.5 VL 28B A3B is Baidu's compact vision-language model, using mixture-of-experts to activate just 3 billion of its 28 billion total parameters. It handles image understanding, multimodal chain-of-thought reasoning, and task-specific fine-tuning via a heterogeneous MoE architecture with separate routing for text and vision modalities. The model aims to deliver strong multimodal performance at modest computational cost, with a 30K token context. It supports both standard and thinking modes for flexible reasoning, and ships as open weights under the Apache 2.0 license with quantized variants. Post-training includes supervised fine-tuning, preference optimization, and reinforcement learning with verifiable rewards for alignment.