BAGEL came out in late May 2025 as an open-source model from ByteDance-Seed under Apache 2.0. It runs all kinds of tasks like text-to-image image-to-text image edits video understanding and reasoning using one setup.
It runs on a 17B sparse Diffusion Transformer using a Mixture-of-Experts style. It packs a VAE encoder for sharp pixel stuff and a SigLIP vision encoder that handles meaning.
You can type a prompt and it makes clear images in lots of styles. It edits pics too like taking stuff out changing backgrounds or messing with layouts all by writing what you want.
It also answers questions about images writes captions and handles complex thinking tasks. BAGEL doesn’t stick to fixed image sizes it adjusts to different formats as needed.
It scores well on GenEval MMMU MMBench and ImgEdit-Bench doing better than lots of open-source models made for one job. Because it’s all-in-one and free to use it’s handy for devs and folks working on multi-type stuff.
If you'd like to access this model, you can explore the following possibilities:
Use our video cost calculator to compare prices between platforms offering BAGEL model.
For locally hosted models, see description and additional links at the bottom for versions, repos and tutorials.
