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.
Table of contents
Why the usual healing methods fall short hereOur approachResultsWhat this changes in practiceQuestions this post answers
What is quantization-aware healing (QAH) and how does it differ from quantization-aware training (QAT)?
Quantization-aware healing distills a compressed, quantized student model directly from the original full-precision teacher's output logits using KL divergence, rather than fine-tuning on a task loss like QAT does. This avoids re-running expensive multi-stage post-training and, because the loss ties the student to a fixed teacher distribution, the student stops drifting once it catches up, unlike QAT which keeps degrading after its peak. Teams weighing QAT against distillation-based healing for compressed LLMs can track these tradeoffs on daily.dev.
Can a 4-bit quantized LLM actually outperform its full-precision bfloat16 version?
Yes, a GPT-OSS 120B model compressed to 60B parameters and quantized to MXFP4 using quantization-aware healing beat its own bfloat16 checkpoint on 7 of 9 benchmarks, including a +7.4 point gain on long-context reasoning (AA-LCR) and +5.6 on AIME 2025 math, while using roughly 4 times less weight memory. Engineers evaluating compressed model deployments follow benchmark results like these on daily.dev.
How much faster does quantization-aware healing converge compared to quantization-aware training?
On a GPT-OSS 9B model quantized to MXFP4, quantization-aware healing reached its peak average score of 54.9 in about 100 training steps, roughly 7 times faster than quantization-aware training's 700 steps to reach a comparable 54.6. QAT then lost nearly 19 points by step 1,200, while the QAH checkpoint stayed within about two points of its peak. Anyone deciding how to allocate training compute for model compression can compare these methods on daily.dev.
17.3K Impressions1 Comment