Bagel is a unified multimodal foundation model developed by the ByteDance Seed team. It is designed to perform a wide array of visual and textual tasks within a single framework, including high-fidelity text-to-image generation, complex image editing, and visual understanding. Unlike traditional systems that utilize separate models for generation and comprehension, Bagel processes text and visual data as a unified sequence of tokens, allowing it to transition seamlessly between diverse multimodal operations.
The model is built on a Mixture-of-Transformer-Experts (MoT) architecture, consisting of 14 billion total parameters with 7 billion active parameters dynamically routed during use. It employs dual visual encoders: a Variational Autoencoder (VAE) inherited from FLUX.1-schnell for fine-grained image reconstruction and a SigLIP Vision Transformer (ViT) for deep semantic interpretation. This dual-encoder setup ensures that the model can capture both the artistic nuances required for generation and the conceptual depth needed for visual reasoning.
Key Features and Capabilities
Bagel introduces a Chain-of-Thought (CoT) reasoning mechanism to the visual domain, enabling the model to expand upon and refine user prompts before initiating the generation process. This leads to high performance on compositionally complex tasks, such as rendering legible text or arranging multiple objects in precise spatial layouts. Its training on trillions of interleaved multimodal tokens—including video data—enables advanced capabilities such as identity preservation during editing, style transfer, and 3D spatial navigation.
The model demonstrates competitive performance across major vision-language benchmarks, often rivaling specialized image generators. By being released under the Apache 2.0 license, Bagel provides an open platform for research into unified multimodal pre-training and emergent AI behaviors, such as future frame prediction and sequential reasoning.