BAGEL

Apache 2.0 licensing is the first thing to know about BAGEL: the weights and code are open, so you can fine-tune, distill and deploy the model wherever you like.

BAGEL is an open-source multimodal AI model that understands, generates and edits both images and text in a single system, and it also handles navigation tasks.

It accepts mixed image and text inputs and returns mixed outputs.

The model is built on a Mixture-of-Transformer-Experts architecture and carries the reasoning and conversational skills of a language model, including multi-turn chat and an optional thinking mode.

For images, it covers editing that keeps visual identity and fine detail, style transfer with minimal alignment data, and in-context abilities such as future frame prediction and 3D manipulation.

Its makers say it matches proprietary systems like GPT-4o and Gemini 2.0 in function and beats other open models on standard benchmarks.

As a research-grade model, you will need to run it yourself.