Why is there no community project for training your own LLM from scratch on consumer hardware?

Posted by tevlon@reddit | LocalLLaMA | View on Reddit | 69 comments

ok so this has been bugging me for a while. We've got nanoGPT/nanoChat from Karpathy which is honestly great and I'd point anyone to it. But here's the thing: to actually follow along and get real results you still end up renting cloud GPUs. And not everyone wants to drop $80+ on cloud compute just to mess around and learn. That barrier alone keeps a ton of curious people out imo. So why isn't there a project (or even just a solid tutorial) built around one hard rule: **it has to train on 8GB of VRAM. no cloud, no rented A100s.** if it doesn't fit on a normal gaming GPU it doesn't count. The dream is a small but actually-real model trained on something like a Wikipedia dump, with a full writeup walking through the whole pipeline. And here's the part I really want: it should use the modern tricks people keep hyping but rarely bundle into one beginner-friendly thing. stuff like: * BitNet / low-bit training to crush the memory footprint * the Muon optimizer instead of plain old AdamW (apparently like 2x more compute efficient + decent memory savings, sounds perfect for a tight VRAM budget) * aggressive quantization to stay inside 8GB * whatever else helps squeeze a trainable model onto consumer hardware basically nanoGPT's vibe but with a hard "must run on your gaming PC" constraint and a modern technique stack, so anyone can train a model end to end for free. so my questions: 1. does this already exist and I just haven't found it? if so please link 2. if not... anyone wanna build it together?