# Qwen3.5-27B-GLM5.1-Distill-v1-4A-4BIT-CHAMPION

**New Predator champion (2026-06-05).** Beats 4A BEST by 0.0056 PPL (chunks=30, Z=22.0, p<1e-19, 30/30 chunks improved).

## Result
- PPL chunks=30: 6.2342 (was 6.2398 for 4A BEST on original F16)
- File size: 16.98 GB (SAME as 4A BEST)
- Same tensor-type-file as 4A BEST
- Different weights (rotated by Chrysalis calibrated for 4-bit)

## How
- Pre-rotated F16 with Chrysalis butterflies calibrated for 4-bit quantization loss (not 2-bit)
- Quantized with PREDATOR_4A_BEST.tensor_types.quant.txt
- 544 matrices × 50 SGD steps × 2.5 min/layer = 158.8 min on VSI 2x L40S

## See also
- 4A BEST (the previous champion, beats Q4_K_M by 0.0369 PPL): Fredred89/Qwen3.5-27B-GLM5.1-Distill-v1-GGUF
- Methodology: 4-bit Chrysalis calibration finds gentler rotations than 2-bit; at 4-bit they HELP (small but real)