PrismML's Bonsai 2 27B Compresses a Full AI Model to 5.9GB With Minimal Loss

On September 17, 2026, AI company PrismML released Bonsai 2 27B, a heavily compressed version of Alibaba's open-source Qwen3.8-27B model that fits into just 5.9GB — down from the original 54GB. The compression uses a ternary weight system, assigning each model weight one of only three values, combined with a mathematical technique called Hadamard rotation to preserve information quality. PrismML claims the compressed model retains 98.2% of the full-precision model's average benchmark score, outperforming conventional 2-bit quantization methods by more than 12 benchmark points while also being smaller in file size. Notably, the model holds up on demanding reasoning tasks like competition mathematics and code generation, areas where standard low-bit compressed models typically degrade significantly. The release is available under the permissive Apache 2.0 license, raising questions about whether high-quality local AI models could reduce reliance on paid subscription services.