Researchers squeezed the 8-billion-parameter Qwen3 language model down to three values per weight, so-called "1.58-bit" compression, and it still holds up surprisingly well.
The pipeline takes a trained language model and squeezes each weight down to one of three values, using a rotation step, an adaptive rounding scheme, and an error-correction pass borrowed from existing quantization research. Applied to Qwen3-8B, the compressed model scored 64.6% on eight benchmark tasks versus 72.4% for the uncompressed version, retaining about 78.5% of its edge over random guessing. Text-prediction accuracy across three test corpora came out 1.36 times worse than the full-precision model. The packed file lands at 8.24 GiB and runs directly in compressed form at 15.52 tokens per second using 7.35 GiB of memory.
The more interesting finding is the comparison, not the compression itself. The same technique applied to a smaller, 4-billion-parameter Qwen3 model kept only 69.6% of its capability, an 8.9-point gap versus the 8B result. That suggests bigger models are simply more forgiving of this kind of aggressive squeezing, worth knowing if you are deciding whether to shrink a large model or just deploy a smaller one from the start.
One catch: the paper's own packed execution kernel, the part meant to make this fast in production, is still slower than standard FP16 GPU code, a reminder that a good compression ratio is not the same thing as a model you can ship.