About QuantDiff

Every popular model gets a dozen GGUF uploads within days of release. They're named the same, sized similarly, and nobody tells you which one is actually closest to the original. QuantDiff runs that diff.

Why "diff"?

A quant is a lossy copy of a model. Like diff on two files, QuantDiff compares the copy with the original, token by token: how far the predicted probabilities drift (KL divergence), how often the top token still matches, whether function calls survive, and what you get back in speed. A small diff means the quant behaves like the real model, and a big diff means you're running something noticeably different.

What we do

For each model up to ~32B parameters, we take the quants people actually download (unsloth, bartowski, mradermacher, and more over time) and diff them against the full-precision model on one fixed machine with fixed settings. We publish the results, the exact commands, the test sets and the raw logs.

Who it's for

People running models locally on 8-16GB consumer GPUs (RTX 3060/4060 Ti/4070/5070/5060 Ti class) who want the best file that fits, rather than the biggest one that technically loads.

The benchmark machine

"Ghost": NVIDIA RTX 5070 12GB, AMD Ryzen 7 5800X, 64GB RAM, Windows. Models up to ~14B run fully on the GPU. 20-32B models run split across GPU and system RAM. That doesn't change their quality scores, and their speeds are clearly labeled as split.

Funding & independence

QuantDiff is funded by GPU-cloud referral links (RunPod, Vast.ai) and clearly marked sponsorships. Sponsors and referral partners never see results early and have no say in which quants rank where. Publishers (model uploaders) cannot pay to be included or excluded. Want to sponsor? Email address coming soon.

Corrections

If you think a result is wrong, the raw logs and settings for that run are on its run page. Re-run it and tell us. Confirmed errors get fixed and noted in the changelog.