Calibrated Qwen3.6-27B quantization tests weight groups before compressing them
enginetown · reddit · 2026-07-28
A solo builder released Qwen3.6-27B-Calibrated after testing which weight groups can be compressed safely before quantization. The project measures divergence one weight group at a time with KL divergence across general, code, math, and tool-use prompts, instead of applying a flat bit-depth everywhere.
The author says the results show a narrow quality cliff around 3.5–3.9 bits per weight group, and that tool calling tends to break first. Three builds were published: Bedrock at 13.26 GB, Tightrope at 12.53 GB, and Gambit at 10.94 GB.
More from Infra
- Meta's commitments jumped to $349B in a quarter as server memory prices tripled ahead of Oct 21 earnings — tengyanAI · 2026-10-16
- Meta releases Muse Spark Open, trained with vLLM-powered RL inference from day one — colinmcnamara · 2026-10-16
- Meta's open Muse Spark needs 493GB of memory — far beyond any consumer laptop — TejasKumar_ · 2026-10-16
- CoreWeave plunges 20% as credit markets price >60% default odds; is this dot-com all over again? — kimmonismus · 2026-10-16
- MiniCPM-V 4.7 35B-A3B hits 17+ tokens/s on iPad with ~3GB memory — techNmak · 2026-10-16
- China rolls out 'token factories' and a 2026-28 token economy action plan — nordicinst · 2026-10-16