Multiverse Computing cuts GPT-OSS to 60 billion parameters and says performance improved

Multiverse Computing researchers have developed a technique called Quantization-Aware Healing that reduced the GPT-OSS model from 120 billion parameters to 60 billion, according to a Techub report citing CryptoBriefing. The team said the compressed version outperformed the full model in seven of nine tests, a result that suggests the reduction did not come at the expense of benchmark performance. The method differs from approaches that rely on an intermediate model during compression. Instead, the compressed model learns directly from the original version. Researchers said the technique can preserve performance while sharply reducing the memory, power, and compute resources needed to run AI models. The report did not provide more detail on the test setup or deployment timeline, but it framed the work as a step toward lowering the operating cost of large models without weakening output quality.