I use a mix of Xcode and Visual Studio for work; mostly mobile, some web dev and API-based AI work mixed in. Up to this point, I've just been copypasta relevant code snippets to GPT for any sort of coding assistance I require. I've experimented with codex web, and I like the way it generates multiple suggestions you can choose from, but I can't integrate it into my workflow yet. Privacy is a concern with some clients.I'm due to refresh my hardware this year, so I'm researching options for local LLMs that I can fit into my workflow more directly. I know that I can run a local model and hook it into VS Code and Xcode, but I currently have no way to really evaluate the available local models and see how practical it would be to run them, as opposed to simply paying for copilot or one of its competitors. I figure $20/mo comes to $720 over three years, which covers the expense of buying a LLM-capable machine. I'm strongly considering Apple's options since they are so cost-effective relative to buying up PC parts right now.Excuse the reddit formatting, here are my questions:1. What local LLMs have you used for coding and how do they compare to something like GPT-5? (yes I know cloud should always be better because more hardware, but I'm talking practical real world differences for coding)2. What is your hardware configuration?3. Since I primarily work off of a laptop, should I just go for a loaded MBP? It doesn't seem to be any more expensive than running a separate Mac Studio/PC as a LLM server. Maybe there is a disadvantage I'm not thinking of.4. From what I can see, simply scaling up model size isn't producing significant performance gains. Assuming that's the case, what would be the most practical RAM amount to buy to stay relatively future proof for a few years?
>$720 for buying a LLM-capable machineI wish. Not in a fat chance in this scalper market.1. Deepseek, GLM2. Three gamer notebooks, common mid range gamer specs. (Budget is a hard limit so is a better bet to use paid llm jewscriptions against that. Note that I get out of memory too).3. Go beyond the brand, into the specs of what each piece of technology you need to run. Then, what those techs require better (arm vs not arm, Nvidia vs AMD, etc). Some models require specifics.4. Best approach is a modular a min rig, enough recent platform that will allow you to do upgrades and limit the model size. Cheapest are the gamer prebuilts.Define your problem first. Without that you will get into a large sum of nevers:What size are the models you want to use? What types? Which models? Do you have access to manuals for each? Requirements?Which type of outputs?How will you grade/measure differences between outputs? Who will be the users / only you? When do you need the minimal rig?
>>107484268> I wish. Not in a fat chance in this scalper market.It's not the cost of the machine outright, it's the difference between a typical base productivity PC/Macbook and one that can run a reasonable amount local LLMs. >Define your problem firstAnon did you speedread or what? I'm asking what people have used and how it worked for them. I don't need "buy X". I'm looking for "I run model X for my daily work in language Y and it is good/terrible compared to GPT/Claude/Gemini".
