Reduce quality? Like quantizing models or context cache