ERNIE 4.5 VL 424B A47B

ERNIE 4.5 VL 424B A47B is Baidu's largest multimodal model, scaling to 424 billion parameters with 47 billion active per token via mixture-of-experts. It targets complex image-text understanding and reasoning across visual content through heterogeneous MoE pre-training that supports both shared and dedicated parameters across textual and visual modalities with modality-isolated routing. It is built to perform strongly across multimodal benchmarks, with thinking mode adding advantage on reasoning tasks such as MathVista, MMMU, and visual puzzles. The model supports a 123K token context for extended document and image analysis. Post-training incorporates supervised fine-tuning, preference optimization, and reinforcement learning with verifiable rewards, and it is released as open weights under the Apache 2.0 license, optimized for inference via vLLM with tensor parallelism across multiple GPUs.