KronQ Pushes 2-Bit Quantization of LLaMA-3-70B to 7.93 Perplexity Where GPTQ Falls Apart

arXiv (COLM 2026) / HuggingFace Daily Papers·medium signal

KronQ (arXiv:2607.07964, submitted July 8, COLM 2026, trending on HuggingFace today) is a post-training quantization method that folds gradient covariance — not just activation statistics — into the pipeline, using bidirectional incoherence processing and a Hessian-trace sensitivity metric for mixed-precision layer allocation. On aggressive 2-bit weight-only quantization of LLaMA-3-70B it reaches 7.93 WikiText-2 perplexity, while GPTQ and GPTAQ diverge into degenerate outputs (>2000 perplexity). For anyone trying to fit a 70B model onto constrained local hardware, this is meaningful evidence that usable 2-bit is getting closer.