I have to say in my experience falcon-40b-instruct got very close to chatgpt (gpt-3. 5),even surpassing it in few domains. However, it is important to note (not at all)OpenAI are doing tricks with the model output. So comparing OS models with just greedy output decoding (very simple) is not fair for OS models.

Still, I'm very excited this model at 13B seems to be matching falcon-40B in some benchmarks. I'm looking forward to using it :-)