ERNIE 4.5 VL 424B A47B

ERNIE 4.5 VL 424B A47B is Baidu's flagship multimodal foundation model, built around a heterogeneous mixture-of-experts architecture that scales to 424 billion total parameters while activating 47 billion per token. Modality-isolated routing allows shared and dedicated expert parameters across textual and visual streams, letting the model tackle complex image-and-text understanding without collapsing the two domains into a single pool. Its long context window supports extended documents and dense visual inputs, making it well-suited to workloads such as multi-page analysis with embedded figures, chart interpretation, and other tasks where reasoning must span both prose and pixels.

Post-training combines supervised fine-tuning, preference optimization, and reinforcement learning with verifiable rewards, and a dedicated thinking mode has been shown to lift performance on multimodal reasoning benchmarks including MathVista, MMMU, and challenging visual puzzles. Because the weights ship under an Apache 2.0 license, teams can self-host at scale using inference paths such as vLLM with tensor parallelism across multiple GPUs. The combination of open weights, a strong multimodal reasoning focus, and tool-calling support positions the model as a practical backbone for applications that need both visual comprehension and careful step-by-step problem solving.