KV cache for dense models is order 50% of parameters. For sparse moe models it can be significantly smaller I believe, but I don’t think it is measured in kb.