So, there are two models that have both been released in the last couple of months which are both DEMONS for tool calling and tasking. First of all the qwen 3.5 family, didnt sleep last night so cant remember the sizes but there are plenty of small models and they're EXCEPTIONAL at everything except general/world knowledge (small size after all), but they rock with tools, code, tasks and so on, as long as you give them a decent context size (this is important, if your context fills then expect the model to start screwing instantly). Second is gemma 4, while not quite as capable in some way, gemma has some really good tool calling skills and is better than qwen at languages.

Both are absolute monsters for their small sizes. Not sure if you're using quants, quantizing your KV cache or whatever but tuning small models can be pretty important. Just want to mention about your context size again though, if you set this to small on agentic tasks then you're gonna hit problems. Its possible(?) that the models you tried didnt have enough context to finish the tasks, you mentioned online AI works fine but online models usually have a MASSIVE context (google is 1M tokens, I think the default for llama.cpp/lmstudio/most stuff is only 8096).

I hope you made some progress since you posted this, but if you can be bothered let me know how things are going, even more so if you try either model I suggested. Happy AI'ing buddy!