Boole-1.7B: A New 1-Bit Qwen3-1.7B Model Outperforming Bonsai-1.7B
Precision Numerics has released Boole-1.7B, an extreme 1-bit quantized version of the Qwen3-1.7B model. Weighing only 247 MB, it achieves a 13.9x size reduction compared to BF16 while outperforming the existing 1-bit Bonsai-1.7B baseline. This release marks a significant milestone in edge AI and local language models by demonstrating that extreme 1-bit quantization can retain high fidelity. Such ultra-lightweight models enable efficient execution on resource-constrained devices, wearables, and directly within web browsers. Boole-1.7B retains 85.9% of the original BF16 model's average score across a 10-task evaluation benchmark. It supports local execution through vLLM, llama.cpp, and can even be tested directly within web browsers.
## BACKGROUND
Quantization in large language models reduces the precision of model weights to decrease memory footprint and accelerate inference speed. 1-bit quantization compresses each parameter down to just a single bit, drastically shrinking model size to fit onto edge devices while requiring specialized software kernels to function.