Features
Unified multimodal understanding and generation in one model
Autoregressive framework with unified Transformer architecture
Decoupled visual encoding pathways (SigLIP-L + MLP adapters)
Text-to-image generation with instruction following
Image understanding (description, analysis)
Available in 1B and 7B parameter variants
MIT license, open-source on Hugging Face and GitHub
WebGPU support for in-browser inference (1B model)
Compatible with ComfyUI via community nodes
Bidirectional image-text interaction
Optimized training with extended datasets and stability techniques
Model sizes: 1.3B (Janus), 1.3B (JanusFlow), 1B, 7B checkpoints
Sequence length 4096 for all models
Integration with Hugging Face and GitHub for deployment
Supports unrestricted commercial use
Official LLM leaderboard and ranking
Battle Mode for side-by-side model comparison
Agent Mode for real-world agent evaluation
Multimodal Max for multimodal model testing
Code Arena with web development challenge categories
BullshitBench for nonsense detection evaluation
File upload for third-party AI processing
Public conversation sharing for community research
Leaderboard dataset publicly available
Monthly arena updates across product and research
Search functionality for models and conversations
Community voting on model responses
