<!-- mobian-agent-page publisher="dailydev" canonical="https://daily.dev/posts/FYfUQBqZR" -->

---
title: Quantization-Aware Healing: a compressed, 4-bit model...
description: Multiverse Computing introduces Quantization-Aware Healing (QAH), a recovery technique for models that have undergone both structural compression and...
canonical: https://daily.dev/posts/quantization-aware-healing-a-compressed-4-bit-model-that-outperforms-its-full-precision-original-fyfuqbqzr
twitter:card: summary_large_image
twitter:site: @dailydotdev
og:type: website
og:site_name: daily.dev
og:title: Quantization-Aware Healing: a compressed, 4-bit model that outperforms its full-precision original | daily.dev
og:description: Multiverse Computing introduces Quantization-Aware Healing (QAH), a recovery technique for models that have undergone both structural compression and...
og:url: https://daily.dev/posts/quantization-aware-healing-a-compressed-4-bit-model-that-outperforms-its-full-precision-original-fyfuqbqzr
og:image: https://api.daily.dev/og/posts/FYfUQBqZR.png
og:image:alt: Quantization-Aware Healing: a compressed, 4-bit model that outperforms its full-precision original
og:image:width: 1200
og:image:height: 630
og:locale: en
---

> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Quantization-Aware Healing: a compressed, 4-bit model that outperforms its full-precision original

**[Hugging Face](https://daily.dev/sources/huggingface)** · 8 min read · 20 upvotes · 1 comments

## Summary

Multiverse Computing introduces Quantization-Aware Healing (QAH), a recovery technique for models that have undergone both structural compression and quantization. Applied to a GPT-OSS 120B model compressed to 60B parameters and quantized to MXFP4, QAH distills directly from the original full-precision teacher rather than from the recovered bfloat16 checkpoint, producing a 4-bit model that beats its own bfloat16 source on 7 of 9 benchmarks and even surpasses the full-size teacher on LiveCodeBench. Compared head-to-head against quantization-aware training (QAT) on a GPT-OSS 9B model, QAH reaches a similar peak accuracy roughly 7 times faster and remains stable afterward, while QAT collapses sharply past its peak. The result uses roughly 4x less weight memory and roughly half the compute per token versus the teacher.

## Full article

daily.dev links to this article rather than hosting it. Read it at the original source: <https://huggingface.co/blog/MultiverseComputingCAI/quantization-aware-healing>

## Questions this post answers

### What is Quantization-Aware Healing (QAH) and how does it differ from quantization-aware training (QAT)?

Quantization-Aware Healing distills a structurally compressed and quantized model directly from the original, pre-compression full-precision model rather than from a recovered checkpoint, using KL divergence on output logits. Unlike QAT, which fine-tunes with a task loss and can drift or degrade after its peak, QAH ties the student to a fixed teacher distribution, reaching peak accuracy in about 100 steps versus QAT's roughly 700 steps, with far greater post-peak stability.

_Teams choosing between healing recipes for compressed models can track distillation techniques like this on daily.dev._

### Can a 4-bit quantized LLM outperform its own full-precision bfloat16 version?

Yes, a GPT-OSS model compressed to 60B parameters and quantized to MXFP4 using Quantization-Aware Healing beat its own bfloat16 checkpoint on 7 of 9 benchmarks, with the biggest gains on long-context reasoning (+7.4 on AA-LCR) and math (+5.6 on AIME 2025), while losing by under 1.5 points on the remaining two.

_Developers evaluating quantized model tradeoffs can follow benchmark results like these on daily.dev._

### How much memory and compute does a QAH-recovered 4-bit model save compared to its full-size teacher?

A QAH-recovered model at 4-bit precision uses roughly 4 times less weight memory than its bfloat16 counterpart, and at half the parameter count of the original full-size teacher it roughly halves compute per token; for model families shipped natively in bfloat16, the combined savings approach 8 times less compute per token.

_Anyone sizing hardware for compressed model deployment can track efficiency gains like these on daily.dev._

## Community discussion

Top comments from developers on daily.dev.

**@paoloap** · 3 upvotes

> watched qat collapse right after peak too, you need early stopping and a babysitter to catch the good checkpoint. healing that stays stable past peak is what makes it usable in a real pipeline.

## Similar posts on daily.dev

- [Quantization-Aware Training in TorchAO \(II\) – PyTorch](https://daily.dev/posts/quantization-aware-training-in-torchao-ii-pytorch-3qafp0fko) · PyTorch · 1 upvotes · 0 comments
- [Accelerating LLM inference with post-training weight and activation using AWQ and GPTQ on Amazon SageMaker AI](https://daily.dev/posts/accelerating-llm-inference-with-post-training-weight-and-activation-using-awq-and-gptq-on-amazon-sag-jbmt9237k) · AWS · 3 upvotes · 0 comments
- [5 LLM Quantization Techniques](https://daily.dev/posts/5-llm-quantization-techniques-nfunrpnuc) · Daily Dose of Data Science \| Avi Chawla \| Substack · 3 upvotes · 1 comments
- [Model Quantization: Concepts, Methods, and Why It Matters](https://daily.dev/posts/model-quantization-concepts-methods-and-why-it-matters-gbx1qm1yl) · NVIDIA Developer · 2 upvotes · 0 comments

---

Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#data-science](https://daily.dev/tags/data-science), [#llm](https://daily.dev/tags/llm)

[View this post on daily.dev](https://daily.dev/posts/quantization-aware-healing-a-compressed-4-bit-model-that-outperforms-its-full-precision-original-fyfuqbqzr)

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]}
{"@context":"https://schema.org","@type":"TechArticle","headline":"Quantization-Aware Healing: a compressed, 4-bit model that outperforms its full-precision original","url":"https://daily.dev/posts/quantization-aware-healing-a-compressed-4-bit-model-that-outperforms-its-full-precision-original-fyfuqbqzr","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/quantization-aware-healing-a-compressed-4-bit-model-that-outperforms-its-full-precision-original-fyfuqbqzr"},"datePublished":"2026-08-25T11:42:56.593Z","dateModified":"2026-09-13T21:24:13.105Z","description":"Multiverse Computing introduces Quantization-Aware Healing (QAH), a recovery technique for models that have undergone both structural compression and...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/aba5f7f6cf091b57e3b220bbd619cddd?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/aba5f7f6cf091b57e3b220bbd619cddd?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hugging Face","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hugging Face","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/f1f55c67d81a4330acf5b90b26b0c8e1","url":"https://daily.dev/sources/huggingface"},"commentCount":1,"discussionUrl":"https://daily.dev/posts/quantization-aware-healing-a-compressed-4-bit-model-that-outperforms-its-full-precision-original-fyfuqbqzr","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":20},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":1}],"keywords":"machine-learning,data-science,llm","timeRequired":"PT8M"}
{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hugging Face","item":"https://daily.dev/sources/huggingface"},{"@type":"ListItem","position":3,"name":"Quantization-Aware Healing: a compressed, 4-bit model that outperforms its full-precision original"}]}
{"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/quantization-aware-healing-a-compressed-4-bit-model-that-outperforms-its-full-precision-original-fyfuqbqzr","comment":[{"@type":"Comment","text":"watched qat collapse right after peak too, you need early stopping and a babysitter to catch the good checkpoint. healing that stays stable past peak is what makes it usable in a real pipeline.","datePublished":"2026-08-25T16:47:45.354Z","url":"https://daily.dev/posts/FYfUQBqZR#c-nhBMGTcsR","author":{"@type":"Person","name":"Paolo Perrone","url":"https://daily.dev/paoloap","image":"https://lh3.googleusercontent.com/a/AGNmyxaOowFqs19Msmo4AtbOanvSwkOb705dhJ-u3fFQ0Nw=s96-c"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":3}}]}
{"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/quantization-aware-healing-a-compressed-4-bit-model-that-outperforms-its-full-precision-original-fyfuqbqzr#faq","mainEntity":[{"@type":"Question","name":"What is Quantization-Aware Healing (QAH) and how does it differ from quantization-aware training (QAT)?","acceptedAnswer":{"@type":"Answer","text":"Quantization-Aware Healing distills a structurally compressed and quantized model directly from the original, pre-compression full-precision model rather than from a recovered checkpoint, using KL divergence on output logits. Unlike QAT, which fine-tunes with a task loss and can drift or degrade after its peak, QAH ties the student to a fixed teacher distribution, reaching peak accuracy in about 100 steps versus QAT's roughly 700 steps, with far greater post-peak stability. Teams choosing between healing recipes for compressed models can track distillation techniques like this on daily.dev."}},{"@type":"Question","name":"Can a 4-bit quantized LLM outperform its own full-precision bfloat16 version?","acceptedAnswer":{"@type":"Answer","text":"Yes, a GPT-OSS model compressed to 60B parameters and quantized to MXFP4 using Quantization-Aware Healing beat its own bfloat16 checkpoint on 7 of 9 benchmarks, with the biggest gains on long-context reasoning (+7.4 on AA-LCR) and math (+5.6 on AIME 2025), while losing by under 1.5 points on the remaining two. Developers evaluating quantized model tradeoffs can follow benchmark results like these on daily.dev."}},{"@type":"Question","name":"How much memory and compute does a QAH-recovered 4-bit model save compared to its full-size teacher?","acceptedAnswer":{"@type":"Answer","text":"A QAH-recovered model at 4-bit precision uses roughly 4 times less weight memory than its bfloat16 counterpart, and at half the parameter count of the original full-size teacher it roughly halves compute per token; for model families shipped natively in bfloat16, the combined savings approach 8 times less compute per token. Anyone sizing hardware for compressed model deployment can track efficiency gains like these on daily.dev."}}]}
```

