4-Bit GPT-OSS 60B Beats Its Own BF16 Checkpoint on 7 of 9 Benchmarks

Multiverse Computing's research team has published a result that inverts one of the field's standing assumptions about model compression: a GPT-OSS 120B model structurally compressed to 60B parameters and then quantized to MXFP4 outperforms its own bfloat16 checkpoint on 7 of 9 benchmarks. The method, called Quantization-Aware Healing (QAH), is described in the paper Quantization-Aware Healing: A Practical Recipe for Recovering Compressed, 4-Bit LLMs (arXiv 2608.20953), published August 25, 2026. For practitioners running compress-then-heal pipelines — the architecture shared by GPT-OSS, NVIDIA's Nemotron family, and Multiverse's own Hypernova 60B — this is a concrete, reproducible departure from the assumption that 4-bit precision always costs accuracy relative to 16-bit.

Why Existing Healing Methods Fail After Structural Compression

The standard three-step pipeline — structural compression, quantization, then healing — leaves the healing step underspecified. Quantization-aware training (QAT) inserts fake-quantization operators into the forward pass and continues fine-tuning on a task loss, meaning the entire post-training sequence (SFT, RLHF, agentic tuning) reruns through a noisier, lower-precision forward pass. Beyond the cost, the paper's results show QAT is unstable past its accuracy peak, shedding nearly 19 points on an average of MMLU-Pro, LiveCodeBench, and GPQA Diamond by step 1,200 after peaking around step 700.

Quantization-aware distillation (QAD) sidesteps retraining cost by distilling a frozen full-precision teacher into the quantized student via KL divergence on output logits. That works when only quantization has occurred, because a matching full-precision model exists. But after structural compression, no independently trained full-precision version of the smaller architecture exists. The only available teacher is the recovered bfloat16 checkpoint, itself a distilled approximation — which caps the quantized student at whatever ceiling the bfloat16 recovery step already imposed.

The QAH Mechanism

QAH resolves the ceiling problem by distilling directly from the original, pre-compression model rather than from the recovered bfloat16 checkpoint. Teacher and student share neither architecture nor parameter count: the teacher is the full 120B model at full precision; the student is 60B and running in MXFP4. Because KL divergence operates on output logits — which are architecture-agnostic — the size mismatch does not impede the transfer.

Under QAH, the 4-bit student receives a second, full pass of distillation against the original teacher — supervision the bfloat16 checkpoint never had. The student is not compensating for precision loss; it is absorbing information the earlier recovery stage did not transfer. A frozen teacher's output distribution creates no pressure for the student to drift once it has matched that distribution, whereas cross-entropy against hard labels keeps pushing weights indefinitely.

In a controlled head-to-head on a GPT-OSS 9B model quantized to MXFP4, QAH reaches its peak average score of 54.9 in roughly 100 steps — approximately 7× faster than QAT's peak of 54.6 at step 700 — and then holds within about two points of that peak for the remainder of training. QAT collapses sharply, shedding nearly 19 points by step 1,200.

For long-context healing, where the corpus includes documents up to 32,768 tokens, QAH reuses a chunked KL-divergence loss from the team's companion distillation paper. The loss computes KL one sequence slice at a time, avoiding materialisation of the full vocabulary-by-sequence tensor to keep 32k-token training within a fixed GPU memory budget.

Benchmark Results

The primary evaluation compares three checkpoints: the 120B teacher quantized to MXFP4, the 60B bfloat16 recovered checkpoint, and the 60B MXFP4 model healed with QAH.

The largest gains appear on exactly the capabilities structural compression damages most: long-context reasoning (+7.4 on AA-LCR) and mathematical reasoning (+5.6 on AIME 2025). The two benchmarks where QAH trails — MMLU-Pro and SciCode — are deficits of 0.2 and 1.4 points respectively. Against the 120B teacher, the QAH model surpasses it on LiveCodeBench (66.5 vs. 66.0) and lands within 1.6 points on GPQA Diamond (67.4 vs. 69.0), despite running at half the parameter count and roughly a quarter of the weight memory.

Deployment Consequences

The efficiency arithmetic stacks multiplicatively. At half the teacher's parameter count, the QAH model roughly halves compute per token; at 4-bit versus bfloat16 precision, weight memory shrinks by approximately 4×. For model families that ship bfloat16 checkpoints rather than already-quantized weights, the combined reduction relative to the full-precision original approaches 8× less compute per token. The QAT stability risk — needing precise early stopping to avoid shipping a degraded checkpoint — disappears under QAH because the frozen-teacher loss provides no gradient signal to continue drifting once the student converges.

This connects to a pattern visible across the efficiency stack: as covered in our analysis of how software extraction is displacing hardware acquisition at the AI frontier, the most consequential gains in 2026 are coming from training and compression methodology, not raw parameter scaling. QAH is a pointed example — the 4-bit model does not merely recover accuracy lost to compression; it exceeds the bfloat16 baseline on seven tasks while simultaneously reducing serving cost.