Janus is a multimodal model developed by DeepSeek, capable of both understanding and generating images. It employs separate encoding pathways optimized for each task within a unified transformer-based architecture. This design allows Janus to accurately extract semantic information from images for comprehension tasks while preserving detail and coherence for generation tasks. With 1.3 billion parameters and a sequence length of 4096, Janus is both compact and efficient.

The team has also developed a Telegram bot that interfaces with their models, enabling users to interact with Janus directly through the Telegram platform.