Measure how much a quantized model drifted from the original.

quantgap is an open-source tool for measuring the distribution drift introduced by
quantization. It scores models locally against a shared reference pack, so different
quant formats and publishers can be compared using the same method.
What quantgap provides
- KLD drift measurements with confidence intervals
- Top-1 agreement and perplexity-ratio metrics
- Checks for looping, wrong-script output, and invalid JSON
- Support for local files and compatible model servers
- Reproducible result files with model and pack hashes
- A public leaderboard for comparing quantized models
Useful links
Code is licensed under Apache-2.0.