A proposal to run part of Claude Code on users’ own computers
Anthropic could offer optional that handle some tasks on the user’s own machine. The main idea is to use the powerful computers many developers already have, instead of sending every step to Anthropic’s servers. The model would not need to be open source, and it could work only inside Claude’s own hook system.
One suggested setup is a 30 billion parameter model, or several smaller 8 billion parameter models that each handle specialized tasks in parallel. This could reduce token use, keep quality similar, and possibly make some agent work faster.
Key points
- Claude Code could offload some to a user’s own computer.
- The goal is to reduce token use and server compute needs.
- A 30 billion parameter model or several 8 billion parameter specialist models were suggested.
- The would not have to be open source.
- The hoped-for result is similar quality with lower cost and possibly faster runs.