BAGEL

BAGEL is a unified generation and understanding model that can handle both image and text inputs and outputs in a mixed format. It is pre-trained from large language models, giving it foundational capabilities for reasoning and conversation, and is designed to be fine-tuned, distilled, and deployed in various settings. The model is presented as open and intended to offer functionality comparable to proprietary systems, while being available in an open form for broad use.

BAGEL is trained on interleaved video and web data, which supports a multimodal chain-of-thought during generation. The architecture includes separate encoders for pixel-level and semantic-level features and follows a next group of token prediction paradigm. This training approach aims to enable natural multimodal reasoning, editing, and multi-turn interactions within a unified interface.

BAGEL outlines a range of capabilities including high-fidelity image and video generation, image editing, style transfer, navigation, and reasoning tasks. The model supports complex tasks such as predicting future frames, 3D manipulation, and world navigation, while maintaining coherent and contextually aware outputs through multimodal understanding.

BAGEL describes its plans and scope as open-source and intended for broad deployment, emphasizing multimodal capabilities across language, image, and video data. The material references integration with existing tools and data modalities, and highlights the scalable, token-based training paradigm designed to improve understanding, generation, and editing tasks across diverse environments.

Use cases

• Multimodal reasoning and generation tasks
• High-fidelity image and video generation
• Image editing with preserved details