PrismML Releases Ternary Bonsai 2 27B 1.58-Bit Language Model
PrismML has released Ternary Bonsai 2 27B, a 27-billion parameter open-weights language model operating with 1.58-bit ternary quantization across its architecture. The GGUF model files are available on Hugging Face for running on local devices and inference engines like vLLM and LM Studio. This release marks an important advancement in extreme LLM compression, allowing a 27B parameter model to fit within significantly reduced memory budgets. By drastically lowering memory bandwidth demands, 1.58-bit models make running large AI models locally far more accessible and energy-efficient. Unlike post-training hybrid quantization approaches, Ternary Bonsai implements full 1.58-bit weight representation throughout the entire model network without relying on higher-precision escape hatches. The GGUF artifacts allow deployment across frameworks such as vLLM, Docker, and LM Studio.
## BACKGROUND
Ternary or 1.58-bit quantization constrains neural network weights to three values: {-1, 0, +1}. Because log2(3) is approximately 1.58, each weight only requires 1.58 bits of storage rather than standard 16-bit floating-point (FP16) values. This encoding enables replacing heavy matrix multiplication operations with simple additions and subtractions, drastically reducing hardware compute and memory overhead.