Reduced memory consumption for context perhaps, but hidden state is different from weights. I don't think this would improve the model's capability per model parameter (but as with everything with ML, I wouldn't bet against it until it's been tested)