The minimum size of an LLM to generate a little meaningful convo

I am trying to find the smallest model to generate something in "normal" English, it could be like AI generated, but it should be mostly grammar correct, sensible in the context. I hope it could be less than 10M to be easy for CPU, so I tested some like Supra-Mini-v6-1M. well it's not good enough. `output // its out put for prompt "Once upon a time": rounded, and dripped a small part and mi, while others. Clicy for example in Papuano Pauder (18159), is often found as "maping with all" in which we believe from God; (2036): “Body's Give to do a sight is in your bust and is not very old." Approaching it may look a few days, which includes any kind which gives you to yourself if we have the best name or other cosy to be heard in it). But that you want it, because they say there should only need some kinds to have all you would want you and then you must use their good good fellow, and I can’ln? (1-5): 1.) What does any good biblo mean and do you know that the pets you have come with or are all those that they love me for a lot to learn and give me the word? This was when ` it's even worse than template generated ones. Which micro LLM can get better result? what's the minimum parameters? Thanks!