Master’s thesis: post-training quantization

Goal

Choosing an AI model today means trusting benchmark charts that never say how the numbers were produced: which hardware, which software, which shortcuts. The thesis builds genuine understanding of what quantization does to models, so those choices stop being guesswork. Done right, that understanding could lead somewhere entirely new.

Report

One motivated student, broad scope, early days: hands-on benchmarking of language models mixed with digging into post-training quantization methods, with explicit room to try novel approaches if the intuition leads there.

No results yet. The exciting part is still ahead.