SINQ
SINQ is a calibration-free quantization method that shrinks large language models to lower precision while keeping most of their accuracy, quantizing a 14B model in about 21 seconds and a 236B model in about 5 minutes.
SINQ, Sinkhorn-Normalized Quantization, is a quantization technique for large language models that requires no calibration data, works with symmetric and asymmetric quantization, and is designed to be applied to any model without per-model tuning. A calibrated variant, A-SINQ, is also provided for cases where calibration data is available.
It integrates with Hugging Face Transformers through a SinqConfig class compatible with AutoModelForCausalLM, so a model can be loaded with SINQ quantization applied directly. A related pre-SINQ algorithm combines SINQ's normalization with GGUF quantization, and the project publishes a growing collection of pre-quantized GGUF models built this way.
SINQ is written in Python and released under the Apache license by Huawei's CSL team; the accompanying paper was accepted at ICML 2026.