Any good local model for subagent?

Posted by AreaExact7824@reddit | LocalLLaMA | View on Reddit | 2 comments

Is 8B model enough for searching related file? Wanna optimize token

[-]

Joozio@reddit

For file search 8B is enough if you give it good filtering scaffolding around it. Qwen 8B handles grep-style 'find files matching X' fine for me. Where it falls over is when you let it pick search strategy on its own. I keep the heuristics in code and let the model just rank or summarize. Saves tokens and the failure modes are smaller.

[-]

Substantial_Step_351@reddit

For file search yeah, 8B is more than enough. It's a narrow enough task that you don't need serious reasoning horsepower, just solid instruction following and clean output. Qwen3 at the 8B range is probably your best bet right now if you want strong tool use out of the box. Phi-4 Mini is worth a look too if you want something leaner

The failure mode I'd actually watch isn't model size, it's prompt ambiguity. If "related" isn't tightly defined, even a good model will drift. Tight matching criteria in the prompt will get you further than bumping up to a 14B