AI BriefingSana 1.6B Model Released with 1.58-bit Quantization
·2026.06.28 14:10
Key point
The Sana 1.6B text-to-image model was quantized to 1.58-bit ternary, reducing its size by 8.6x.
Details
Clark Labs has released a ternary (approximately 1.85 bits/weight) quantized version of the Sana 1.6B text-to-image transformer model.
Key features are as follows:
- Overwhelming compression ratio: Compared to the FP16 model (3.21 GB), the size was reduced to about 12% (374 MB), making it 8.6x lighter.
- Quality preservation: The model maintains quality similar to FP16 even after quantization, with some layers requiring precision (about 5%) designed to retain higher precision.
- Compatibility: The transformer/dequantized bf16 model in this repository is immediately compatible with the diffusers library.
This model dramatically reduces model size while preserving image generation quality, making it optimized for deployment and inference in resource-constrained environments.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.