PrismML has introduced Ternary Bonsai 2 27B, a compressed version of Alibaba's Qwen3.8 27B language model. According to MarkTechPost, the model reduces the original size from 53.80 GB to 5.93 GB by quantizing every weight to one of three values: -1, 0, or +1. This ternary approach drastically cuts memory requirements while aiming to preserve performance.
The technique, known as ternary quantization, represents weights with just two bits, enabling significant compression. PrismML's implementation is part of a growing trend to make large language models more accessible on consumer hardware. The company has not yet released detailed benchmarks, but the size reduction alone marks a notable engineering achievement.
While the exact performance trade-offs remain unclear, the development highlights ongoing efforts to optimize AI models for edge devices and lower-resource environments. Further details are expected from PrismML or independent evaluations.
