Baidu: ERNIE 4.5 VL 28B A3B

Processes both text and image inputs simultaneously using a 28B parameter Mixture-of-Experts architecture where only 3B parameters activate per token. Implements modality-isolated routing, meaning separate expert pathways handle text and vision features before fusion, enabling specialized processing for each modality without forcing them through identical computational paths. This heterogeneous MoE design allows the model to maintain distinct reasoning chains for language and vision while sharing a unified token-level gating mechanism.

Answers natural language questions about image content by grounding language understanding in visual features extracted through the vision expert pathway. The model performs token-level fusion of image embeddings and text tokens, allowing it to generate answers that reference specific visual regions or objects mentioned in questions. This capability leverages the modality-isolated routing to maintain separate visual reasoning before integrating with language generation.

Analyzes documents, forms, and screenshots by simultaneously processing visual layout and text content through separate expert pathways that fuse at the token level. The model can extract structured information from documents (tables, forms, receipts) by understanding both the spatial arrangement of elements (vision pathway) and semantic meaning of text (text pathway). The heterogeneous MoE architecture allows it to specialize in document structure recognition without diluting text understanding capacity.

Generates natural language descriptions and captions for images by processing visual features through the vision expert pathway and generating coherent text through the text expert pathway with token-level fusion. The model can produce captions at varying levels of detail (short captions, detailed descriptions, technical analysis) based on prompt instructions. The sparse activation pattern (3B/28B) allows efficient batch processing of image captioning tasks.

Maintains multi-turn conversations where users can reference previously shared images and ask follow-up questions that build on earlier visual context. The model preserves image embeddings and visual understanding across conversation turns, allowing users to ask 'what was in that image from earlier?' or refine questions about previously analyzed images. The heterogeneous MoE routing maintains separate visual and text reasoning chains that can be reused across turns without reprocessing images.

Performs reasoning tasks that require simultaneous understanding of both text and visual semantics, such as determining if an image matches a text description, identifying contradictions between image content and text claims, or reasoning about abstract relationships between visual and textual information. The modality-isolated expert routing allows the model to develop independent semantic representations in each modality before fusion, enabling more nuanced cross-modal reasoning than models that force both modalities through identical pathways.

Processes multiple image-text pairs or sequential multimodal requests efficiently through sparse MoE activation, where only 3B of 28B parameters activate per token. This enables higher throughput and lower latency for batch operations compared to dense models, making it suitable for processing large volumes of images with associated queries. The sparse activation pattern reduces memory footprint and computational cost per request, allowing more concurrent requests on the same hardware.