These Researchers Just Shrunk an AI Model and Somehow Made It Smarter
Researchers built a compressed 60B-parameter version of OpenAI’s GPT-OSS 120B that performed better than the normal 60B “half-shrunk” model on most benchmarks. The method, called Quantization-Aware Healing, differs from standard compression because it teaches the smaller model to match the original full-precision 120B model, rather than an already-degraded intermediate model. That change appears to let quantization act as a benefit instead of a loss. The healed model uses about half the parameters and roughly a quarter of the memory of the original, making it cheaper to run and more practical for smaller labs or local deployment. The team released the model as open weights, though the shrink tool remains proprietary and results were tested only on GPT-OSS.
