Multiverse Computing’s team published a method called Quantization-Aware Healing on the Hugging Face blog on August 25. The researchers compressed OpenAI’s GPT-OSS model from 120 billion to 60 billion parameters using 4-bit quantization. The shrunken version outperformed the full original model on 7 out of 9 tests, a performance considered unexpected in the field. The technique involves training the reduced model directly from the uncompressed original rather than the traditional intermediate copy. The resulting model, named Hypernova-60B and released as open weights on Hugging Face, requires roughly four times less memory while matching or exceeding the original’s performance on most benchmarks.
Source: Read the original article
