PrismML team fit a 27B model into 5.9 GB.
We uncensored it without changing a single weight. ๐ณ
Introducing OrcaRouter Ternary Bonsai 2 27B Uncensored.
Traditional abliteration modifies weights. On a ~1.72-bit ternary model, that means re-quantization โ potentially destroying the quality preserved by QAT.
So we moved abliteration into the runtime.
โ 0 weights modified
โ 0 re-quantization
โ original 5.9 GB pack stays bit-identical
โ 129 residual intervention sites
โ adjustable at inference
โ runs locally on Apple Silicon
No modified checkpoint. Bring the original Bonsai pack + our runtime.
Open source: