Hypernova-60B beats bfloat16 on 7 of 9 benchmarks

A new distillation method beats bfloat16 accuracy on seven of nine benchmarks, per the Quantization-Aware Healing paper.

The claim matters for teams quantizing large models to 4-bit precision, a standard technique for cutting memory and inference cost. Training the same student against the original weights instead reaches comparable accuracy, about 7x faster than matched quantization-aware training.

Most quantization pipelines skip that distinction, distilling the 4-bit student from a structurally compressed model's bfloat16 checkpoint instead of the original weights. That checkpoint is already a distillation-recovered approximation, so the student inherits whatever error crept in earlier.

The resulting model, released open-weight as Hypernova-60B, uses about 4x less weight memory than the teacher and half its parameter count.

Teams distilling 4-bit models from a bfloat16 teacher now have a specific baseline to test against, the original weights themselves.

The paper doesn't say which two benchmarks fell short of bfloat16, or by how much.