Three values per weight: -1, 0 or +1. That's the trick — Qwen-27B goes 54 GB → 5.9 GB and keeps 98% of its benchmark scores. What it loses, and what it takes to run: https://pub.towardsai.net/someone-shrank-qwen-27b-by-9x-and-it-kept-98-of-its-benchmark-scores-7c540dde7a2e
Someone Shrank Qwen-27B by 9x and It Kept 98% of Its Benchmark Scores
Every weight is now -1, 0 or +1, which takes Qwen 3.8 27B from 54 GB to 5.9 GB. What it keeps, what it loses, and what it takes to run it.
Medium
