SINQ (Sinkhorn-Normalized Quantization) is a calibration-free low-precision quantization method for LLMs. It uses optimal transport theory and Sinkhorn normalization to quantize model weights while maintaining model quality. The method is model-agnostic and works with various LLMs including Qwen and DeepSeek models. SINQ has been accepted at ICML 2026 and is natively integrated into HuggingFace Transformers via SinqConfig, enabling simplified deployment of quantized models.