Optimizing LLM for autocorrection for iOS custom keyboard

We are looking for someone to optimize existing llm solutions, or create one from scratch, that we can use for autocorrection within the custom iOS keybaord extension we are creating. This model has to adhere to the following constraints:
- be converted to CoreML
- have a latency of < 50ms
- take up less than 50MB of RAM to run

Full details here: https://www.notion.so/slate-hq/Autocorrect-Master-Doc-f3176e9407964041b5f827b7ed0836e6?pvs=4