Emu

Emu is the official open-source repository for the Emu series of generative multimodal models from BAAI (the Beijing Academy of Artificial Intelligence). The models treat images and text in one unified sequence and are trained with a generative pretraining objective, giving them strong in-context learning across vision-and-language tasks: they can generate images from text, describe and reason about images, and learn new tasks from a few examples in the prompt. The repository has around 1,800 stars and was last updated in 2026.

Overview

The Emu line (Emu, Emu2 and successors) comes from the same lab behind the BGE embeddings and FlagEmbedding projects. Rather than bolting a vision encoder onto a language model, Emu discretises visual input into tokens and trains one autoregressive model over the mixed stream, then instruction-tunes it into a multimodal generalist. The repo collects model code, checkpoints and examples for the series.

Key Features

• Unified multimodal generation: text-to-image, image captioning and visual question answering in one model
• Strong multimodal in-context learning from few-shot examples
• Generative pretraining over interleaved image-text data at scale
• Instruction-tuned variants for chat-style multimodal interaction
• Model weights and inference code released for research use
• From BAAI, the team behind BGE and other widely used open models

💳 Pricing

Free. The Emu code is open source under the Apache-2.0 license and the model weights are released for research; you run them on your own hardware.

Who Is It For

Multimodal AI researchers, engineers prototyping vision-language applications, and anyone comparing open foundation models that handle both image generation and understanding.