QuantizeLab

Stop Waiting for GGUFs

Every week, a research lab drops an exciting new open-weights model. The benchmarks are incredible. And then the entire community does the same dance: we wait for someone with a big enough GPU to manually quantize it into a GGUF, so the rest of us can actually run it on our laptops.

I did that dance one too many times. I don't own a 24GB graphics card, so the DIY route meant cloning llama.cpp, compiling C++ from source, fighting CUDA versions and Python virtual environments, and praying my machine wouldn't OOM. All of that, just to produce one runnable file from weights that are freely available. It's a two-minute task stretched into an afternoon. And if the model you want isn't in the top few hyped releases, you wait days for a file that may never come at all.

So I built QuantizeLab.

The loop is simple: paste a Hugging Face model URL, a serverless GPU worker downloads the raw weights, converts them to GGUF, and pushes the finished file straight back into your own Hugging Face namespace. No local setup, no subscription, no waiting on strangers. Jobs are credit-based, priced by model size, and if a job fails the credits are refunded automatically.

The principle I refused to compromise on is IP sovereignty. Your weights stay your weights. The platform never retains, resells, or gates your models. You keep full ownership of everything you quantize.

It's a solo, self-funded project. Stack: React + TanStack Start on Cloudflare, Supabase for auth and state, and Modal for the serverless GPU pipeline. A real 32B quantization job (Qwen2.5-32B, Q4_K_M) runs in about 20 minutes and costs me about $0.65 in GPU time. That margin is what lets me keep pricing honest: small 1.1B models cost 5 credits, and new accounts get 10 free credits, enough to convert your first small model on the house.

Today it supports GGUF (Q4_K_M) with the rest of the format landscape on the roadmap. There's a live status page, a gallery of finished quantizations, and two engineering posts on the blog about the format tradeoffs.

I built this because waiting for other people's uploads is a solved-problem-shaped hole in the open-weights ecosystem. I'd rather ship the pipeline than wait for someone else to ship the file.

Would love feedback: is the credit-based model the right call for this market, and what's the first thing you'd want quantized that currently doesn't exist?