Hacker News Reader

It's not as though the task requires a frontier size LLM. A small on device model should suffice for most clean up.

But if you did want to run a full size model deepseek v4 flash is so cheap that I doubt even many hours of web browsing would have a noticeable cost.

latency would be the killer. a very tiny delay between pressing a key and hearing a response can make a screen reader feel unresponsive.

A relatively small and fast LLM (30b MoE) would probably suffice. It‘s something you can run locally with a nice GPU.