Quantization-Aware Healing: A Practical Recipe for Recovering Compressed, 4-Bit LLMs

Hugging Face Daily Papers Papers

Summary

Quantization-Aware Healing is a method that recovers compressed 4-bit language models by distilling directly from the original uncompressed model, offering faster and more stable performance than Quantization-Aware Training.

Serving large language models cheaply increasingly means shipping models that are both structurally compressed to a fraction of their parameters and quantized to 4 bits. Together these steps degrade reasoning, mathematics, coding, and long-context behavior enough to require a recovery, or healing, stage before deployment. The default recipe, quantization-aware training (QAT), re-fits the compressed, quantized model to hard labels; in our pipeline it converged slowly and collapsed past its peak. We adopted Quantization-Aware Healing (QAH) instead. Because a structurally compressed model is never independently trained at full precision, its bfloat16 checkpoint is a distillation-recovered approximation of the original; QAH distills the 4-bit student directly from the original, uncompressed model. On a GPT-OSS 120B to 60B to MXFP4 pipeline, the QAH student matches or beats its bfloat16 source on 7 of 9 benchmarks at roughly 4 times less weight memory and half the teacher's parameter count, and is released open-weight as Hypernova-60B. Against a matched QAT baseline it reaches a comparable peak about 7 times faster and stays stable under continued training, without hand-tuned early stopping. We also report deployment lessons, including a large, reproducible quality gap between distributed-training backends. Our aim is a recipe deployable without a multi-week hyper-parameter search.
Original Article
View Cached Full Text

Cached at: 08/25/26, 12:35 PM

Paper page - Quantization-Aware Healing: A Practical Recipe for Recovering Compressed, 4-Bit LLMs

Source: https://huggingface.co/papers/2608.20953

Abstract

Quantization-aware healing recovers compressed 4-bit language models faster and more stably than quantization-aware training by distilling directly from the original uncompressed model.

Serving large language models cheaply increasingly means shipping models that are both structurally compressed to a fraction of their parameters and quantized to 4 bits. Together these steps degrade reasoning, mathematics, coding, and long-context behavior enough to require a recovery, or healing, stage before deployment. The default recipe,quantization-aware training(QAT), re-fits the compressed, quantized model to hard labels; in our pipeline it converged slowly and collapsed past its peak. We adoptedQuantization-Aware Healing(QAH) instead. Because a structurally compressed model is never independently trained at full precision, itsbfloat16checkpoint is adistillation-recovered approximation of the original; QAH distills the 4-bit student directly from the original, uncompressed model. On aGPT-OSS120B to 60B toMXFP4pipeline, the QAH student matches or beats itsbfloat16source on 7 of 9 benchmarks at roughly 4 times less weight memory and half the teacher’s parameter count, and is released open-weight asHypernova-60B. Against a matched QAT baseline it reaches a comparable peak about 7 times faster and stays stable under continued training, without hand-tuned early stopping. We also report deployment lessons, including a large, reproducible quality gap between distributed-training backends. Our aim is a recipe deployable without a multi-week hyper-parameter search.

View arXiv pageView PDFAdd to collection

Get this paper in your agent:

hf papers read 2608\.20953

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2608.20953 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2608.20953 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2608.20953 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles