A research team at Multiverse Computing — a quantum-computing company headquartered in San Sebastián, Spain, which previously released a quantum-inspired compression tool for large language models called CompactifAI — has published a method called Quantization-Aware Healing on the Hugging Face blog on August 25, describing how they compressed OpenAI's open GPT-OSS model from 120 billion parameters to 60 billion and reduced its memory to 4-bit precision—with the smaller model outperforming the full-quality model it was derived from on 7 of 9 tests.

The key to the result: the shrunken model was taught from the original high-quality version, not from a weak halfway copy.

The researchers built a smaller, cheaper version of a large AI model, and the smaller version turned out to be smarter than the model it was shrunk from. Ordinarily, that should not happen—it is like losing muscle and getting stronger at the same time. But the group at Multiverse Computing says it did, and the reason suggests the industry has been shrinking AI models the wrong way.

"For practitioners, the practical message is that in a distillation-based healing pipeline the quantization step is not a cost to be minimized but an additional opportunity for teacher supervision, yielding a model that is simultaneously cheaper to serve, lighter in memory, and at least as accurate as its full-precision counterpart," the researchers wrote in a paper published Friday.

How model compression normally works

An AI model's parameters can be thought of as a giant wall of knobs—numbers that hold everything the model has learned. More knobs generally mean a smarter model, but one that is heavier to run. OpenAI's GPT-OSS 120B has 120 billion of those knobs, and each one costs memory and electricity. GPT-OSS is the model family OpenAI released in August 2025 as its first open-weight language models since GPT-2 in 2019—which is what let outside teams like this one download and modify it.

To ship AI cheaply, companies strip knobs and shrink the survivors. The process is like zipping a photo: the file gets smaller, but too much compression blurs the image, much like going from 4K to 720p. Shrink a model too hard and its performance degrades. That trade-off has been widely accepted.

These researchers went further. They cut GPT-OSS to 60 billion parameters and squeezed each one into a tiny 4-bit slot—the digital equivalent of extreme compression. Normally that would gut the model, but the researchers found a way to actually enhance it. In almost all the benchmarks used for comparison, the smaller model outperformed a model built with full precision.

The photocopy mistake

When you shrink a model, you usually fix its mistakes by comparing it to the "half-shrunk" version—the one with 60 billion knobs and higher precision. That teacher-and-student routine is known in the field as knowledge distillation, a technique formalized in influential work by Geoffrey Hinton and colleagues. The problem is that the halfway model is already a blurry copy of the original. The tiny model is therefore taught to imitate a flawed twin, and it can never outgrow that twin.

What the researchers call "Quantization-Aware Healing" changes the target. It points the small model back at the big, uncompressed original and instructs it to copy that model's answers. The little model learns from the master, not the muddy middle. On 7 of 9 tests, the 4-bit 60-billion model beat the 60-billion twin that was supposed to be its better half. The true 120-billion-parameter model still wins most rounds—size has not been abolished—but the "diet" version cleared a bar nobody thought it could.

Smaller and smarter matters because AI eats hardware. The healed model needs roughly a quarter of the memory and half the knobs of the original. That is the gap between an AI that lives in a data center and one that fits on a decent desktop—or, eventually, your phone. Four-bit quantization is already the workhorse format for local-AI tools such as llama.cpp and Ollama, which run compressed models on consumer hardware—so a method that keeps accuracy at 4 bits lands on a path local developers are already using. A model that can produce better results while consuming half the energy means a great deal for small labs and local developers.

The model is also free. The team released the healed Hypernova-60B model as open weights on Hugging Face, so anyone can download and run it. That fits a running streak of open models clearing surprising bars: a mysterious free model, Ox Alpha, recently beat a Claude system with no known builder behind it; there is the hype around Alibaba's Qwen 3.8 Flash Next; and there is the wave of finetunes that improve small models using reasoning traces from bigger LLMs like Fable or Claude Opus.

Don't over-rotate, though. The shrink tool that makes the 60B student is proprietary, so the recipe is not fully open yet, and the team only tested GPT-OSS—not the Llama, Qwen, or Mistral families.