In brief
- Multiverse Computing's team published a method called Quantization-Aware Healing on the Hugging Face blog on August 25.
- They shrank OpenAI's open GPT-OSS model from 120 billion parameters to 60 billion and compressed its memory to 4-bit—and the small version beat the full-quality model it was copied from on 7 of 9 tests.
- The trick: teach the shrunken model from the original smart version, not the weak halfway copy.
gd2md-html alert: inline image link in generated source and store images to your server. NOTE: Images in exported zip file from Google Docs may not appear in the same order as they do in your doc. Please check the images!
-----> A team of researchers just built a smaller, cheaper version of a big AI model. The smaller one turned out smarter than the version it was shrunk from.
That shouldn't happen. It's like losing muscle and getting stronger at the same time. But a group at Multiverse Computing says it did, and the reason says something about how we've been shrinking AI all wrong.
“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.
Think of an AI model’s parameters as a giant wall of knobs—numbers that hold everything it learned. More knobs, smarter model, but heavier to run. OpenAI's GPT-OSS 120B has 120 billion of those knobs. Each one of them costs memory and electricity.
To ship AI cheaply, companies strip knobs and shrink the survivors. It's like zipping a photo: smaller file, but too much compression and the image blurs (like going from 4K to 720p). Shrink a model too hard and it gets dumb. Everyone accepted that trade.
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 these researchers found a way to actually enhance it.
In almost all the benchmarks used for comparison, the smaller model outperformed a model that was 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. The problem is the halfway model is already a blurry copy of the original. So you're teaching the tiny model to *imitate *a flawed twin. It can never outgrow the twin.
What these researchers refer to as “Quantization-Aware Healing” changes the target. It points the small model back at the big, uncompressed original and says: copy this one'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 still wins most rounds—size hasn't been abolished—but the "diet" version cleared a bar nobody thought it could.
Smaller and smarter is a big deal because AI eats hardware. The healed model needs roughly a quarter of the memory and half the knobs of the original. That's the gap between an AI that lives in a data center and one that fits on a decent desktop—or, eventually, your phone.
So a model that can produce better results while consuming half the energy means a lot for small labs and local developers.
It's 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, the hype around Alibaba’s Qwen 3.8 Flash Next, and 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 isn't fully open yet, and the team only tested GPT-OSS—not the Llama, Qwen, or Mistral families.