A recent paper used a single ternary "trit" ~1.5 bits per parameter for training. https://news.ycombinator.com/item?id=39535800 They said it would be difficult to apply this technique to pre-trained models and had to be trained in 1-trit from scratch.