Wan2.1 + DFloat11 Enables High-quality Text-to-video With 24GB VRAM

LeanQuant Tianyi Zhang

💡 Want to run Wan2.1 text-to-video generation on a single 24GB GPU?

Thanks to DFloat11, now you can!

We're excited to release the DFloat11-compressed version of Wan2.1-14B, a powerful diffusion transformer for generating high-quality videos from text prompts.

Key features:

• 📦 Size reduction: 28.6 GB → 19.4 GB
• 🔁 Same output: Bit-for-bit identical to the original
• ⚡ Minimal slowdown: Only 2.5% slower than the original model

With DFloat11, this 14B parameter model fits comfortably on a single 24GB GPU. There’s no quality loss and no compromise in accuracy.

🔗 Try it out:

• Guide: github.com/LeanModels/DFloat11/tree/examples/examples/wan2.1
• Model: huggingface.co/DFloat11/Wan2.1-T2V-14B-Diffusers-DF11
• Research Paper: arxiv.org/abs/2504.11651