LocalLM Agents? vs Slow pool agent?

since the slow pool is becoming pretty misrable and making me question if I want to keep paying for cursor…

I’d really like to see how other LLMs that I’m able to host locally would fair. Has anyone set up a Local agent? This also seems like it would be a pretty good answer to cursor and claude’s lack of silicon to throw at it. I’m pretty sure a huge chunk of the people using the slow pool could run a local LM for development, especially with nvidia’s DIGITS releasing soon

futureproofing?

I could more than likely run something locally; it would be great if cursor had a way to ‘size’ the right model for your hardware and have a ‘local mode’.

from what i understand theres an issue with running a local model because the ‘shadow codebase’ or whatever they call it is saved on cursor’s servers.

If thats true the model you use needs to be publicly visable so that cursor’s servers can connect to it. I could be totally wrong on this. I only have 12gb vram but I’d prefer to run quantized qwen-2.5-32b-instruct or llama3.2-1b-instruct or one of the nemotron models locally if it meant getting rid of the slow queue time for agents.

I dont mind switching over to cursor-small or even a less capable model but they are only available in chat from what I gather, no agent. Agent does a much better job at picking it’s own context, and chat really only sees what you give it context on, which is fine for starting from scratch but if you’re using cursor to understand or edit a complicated codebase its pretty much useless