BAGEL

BAGEL comes from ByteDance's Seed team as an open source foundation model that unifies multimodal understanding and generation in one network. The architecture is a Mixture-of-Transformer-Experts with 7B active parameters out of 14B total, trained on interleaved text, image, video, and web data, and a single checkpoint handles visual question answering, text-to-image generation, free-form image editing, multiview synthesis, and world-navigation style tasks. The team reports understanding and editing benchmark results competitive with leading open and proprietary systems. Running it locally means a conda environment with Python 3.10, flash-attn, and weights pulled from Hugging Face; full precision wants 32 GB or more of VRAM, while the recommended NF4 quantization brings requirements down into the 12 to 32 GB range. Both the code and the BAGEL-7B-MoT weights are Apache-2.0 licensed, so commercial use is allowed. A hosted demo and project site live at bagel-ai.org, training and evaluation guides ship in the repository, and the project has passed 6,000 GitHub stars.