TermKey7269

Really appreciate this, thanksss!!!

Still you won't be able to get that good coding performance, unless and untill, the model is reasoning with code fine tuned(and still token gen rate will be low). Also giving the model access to inspiration from github or web will increase its performance anyways! so ya I'm not exactly talking about running only 2B model but rather a software that lets you run model(you like), giving it access to web and github and boom, making it better!

yeah I'm talking about same lol, websearch + github search = 🔥just with a smaller highly tuned model so that anyone can run it!

Actually I already used that model( I'm an AI enthusiast so regularly benchmark random models) to code locally with cline and it worked exceptionally well, just the ui wasn't that good, rest everything was working fine

thanks bro but tbh I don't think so, bcz newer high density models like Qwen 3.5 are good at coding and following instructions well without halucinating but ya are not good on ui and complex tasks, which work even better with referance. Also we can rather use a large llm with AirLLM but that will increase time significantly( and I don't want that). And to stop it from halucinating, we can implement the project in steps. Also multiple agent idea is good but then most of the low end gpus don't have that much of vram to run multiple instances in parallel so I will make it run in series. I just want a llm to code great on a low end gpu locally without much runtime : )

what about using Qwen3.5-2B-Claude-4.6-Opus-Reasoning-Distilled-GGUF model because it's quite good at making edits and fixing errors but not that good with full stack dev and ui implementations. Also I think I can use AirLLM to use a higher param model but that will increase the run time significantly

Everything seemed hard untill it was achieved so ya I believe that we can implement that easily, than it looks

why?😭😭