DFloat11/OmniGen2-mllm-DF11

DFloat11 Compressed Model: OmniGen2/OmniGen2 MLLM

This is a DFloat11 losslessly compressed version of the original OmniGen2/OmniGen2 model. It reduces model size by 32% compared to the original BFloat16 model, while maintaining bit-identical outputs and supporting efficient GPU inference.

📊 Performance Comparison

Metric	OmniGen2 (BFloat16)	OmniGen2 (DFloat11)
Model Size	16.23 GB	11.11 GB
Peak GPU Memory (1024×1024 image generation)	18.41 GB	14.36 GB
Generation Time (A100 GPU)	25 seconds	27 seconds

🔧 How to Use

A complete usage guide is available in our GitHub repository (forked from the official OmniGen2 repository).

👉 https://github.com/LeanModels/OmniGen2-DFloat11 👈

🔍 How It Works

We apply Huffman coding to losslessly compress the exponent bits of BFloat16 model weights, which are highly compressible (their 8 bits carry only ~2.6 bits of actual information). To enable fast inference, we implement a highly efficient CUDA kernel that performs on-the-fly weight decompression directly on the GPU.

The result is a model that is ~32% smaller, delivers bit-identical outputs, and achieves performance comparable to the original BFloat16 model.

Learn more in our research paper.

📄 Learn More

• Paper: 70% Size, 100% Accuracy: Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float
• GitHub: https://github.com/LeanModels/DFloat11
• HuggingFace: https://huggingface.co/DFloat11