I’ve been obsessed with Andrej Karpathy’s autoresearch, the idea is that you give an AI agent a training setup and let it experiment with architectures and hyper-params autonomously overnight.
The catch? It’s pretty heavy on VRAM. If you’re like me and don't have an H100 in your bedroom, you probably hit an "Out of Memory" error immediately.
I spent the last few days building Litesearch to fix that. It’s a fork optimized for consumer hardware (tested everything from an old and dusty GTX 970 (I had to reinstall the drivers) to an RTX 3060).
What I added to make it "Lite":
Auto VRAM Scaling (probably the best thing so you dont have to manage everything) : It detects your GPU and automatically shrinks or grows the model to fit. No more manual math to avoid crashes.
Dark Mode GUI: I used customtkinter to build a dashboard so you can actually watch the VRAM bar and training logs in real-time.
The "Try" Button: Once the agent finds a good model, you can chat with it directly in a popup window. It’s cool to see the tiny model actually start to make sense.
Pascal Support: Added an fp32 fallback for those older 10-series cards that don't like bfloat16.
Why am I sharing this?
Because local AI research shouldn't be limited to people with $40k rigs. If you’ve got an old gaming laptop or a mid-range PC, you should be able to run these experiments too.
Check out the repo here: https://github.com/jlippp/litesearch
It’s easy to set up with uv or pip. If you give it a spin, let me know if it works on your specific GPU I'm trying to map out as many compatible GPUs as possible! (okay maybe not a 256MB VRAM card)
This project is still getting updated so make sure to follow it
Top comments (0)