Multiverse says 4-bit compressed GPT-OSS model matches or beats its 16-bit copy on seven of nine tests

Multiverse Computing says a GPT-OSS 120B model that it shrank to 60 billion parameters and then quantized to 4 bits (MXFP4) scores as well as or better than its own 16-bit (bfloat16) version on seven of nine benchmarks. The claim comes from a company paper describing a recovery method called Quantization-Aware Healing, or QAH. The 16-bit model in that comparison is the compressed 60B checkpoint, not the original 120B model.

What is new is the choice of teacher. Compressed models usually need extra training, called healing, to win back lost accuracy. Multiverse argues that a compressed model's 16-bit checkpoint is itself only an approximation of the original, so QAH trains the 4-bit model to imitate the original, uncompressed model directly. On a smaller test that compressed a 20B model to 9B, the authors report that QAH reached a peak comparable to standard quantization-aware training in about 100 steps instead of roughly 700, and that the standard method lost nearly 19 points when training continued to step 1,200.

On the headline 60B model, the paper reports gains of 7.4 points on the AA-LCR long-context test, 5.6 on AIME 2025 and 2.7 on Aider, while MMLU-Pro and SciCode slipped by 0.2 and 1.4 points. Against the original 120B model the picture is less rosy. The 4-bit model matches it on LiveCodeBench (66.5 versus 66.0), which the authors call a tie within run-to-run noise, and it trails by 7.3 points on AA-LCR. The authors estimate roughly four times less weight memory than the 16-bit student.

All of this is Multiverse's own measurement, and this article rests on a single source, the paper as presented on Hugging Face. Guth News found no independent replication. The authors list their own limits: every number is a single run without confidence intervals, some benchmarks are small, only one model family and format were tested, and the compression operator is proprietary. They also did not run the direct test of distilling from the 16-bit checkpoint, so the claim that this approach caps accuracy is, in their words, not directly measured. The released Hypernova-60B model received further training, so these figures are not that checkpoint's published results.

If the recipe holds beyond this setup, the usual accuracy cost of 4-bit serving may depend partly on how the recovery is done. For now it is a promising company result on one model family, not a demonstrated general rule.

Revised 16 October 2026: Guth re-read the original sources and rewrote this article. Correction: the earlier version said the healed model beat "its own full-precision version" and that the source gave no benchmark names or scores. The source says the comparison is with the model's own recovered 60B bfloat16 checkpoint (the original 120B model is only matched on LiveCodeBench), and it names nine benchmarks with scores. The earlier version also described a Multiverse reply about training data that does not appear in the source, and that passage has been removed. The earlier version remains available as revision 1.