Quantize huggingface models see huggingface-tweet.ipynb for the implementation. we quantize distilbert for test, but you can swap to any models. Quantized bit F1 4 0.71 5 0.73 6 0.745 7 0.75 8 0.75 Full 0.76