fabq-rc

FABQ-RC: Fisher-Adaptive Binary Quantization with Residual Codebooks

A Vigorous Scientific Research Experiment

Status: Active Duration: April 2026 - Present

---

What Is FABQ-RC?

FABQ-RC is a 1-bit quantization method for large language models that adapts per layer rather than using a fixed blocksize. It combines:

1. Fisher-Weighted Channel Importance — Which channels actually matter for loss?
2. Mixed-Precision Core Allocation — int8 for critical channels, binary for the rest
3. Adaptive Blocksize — Per-layer blocksize selection, not global
4. Residual Codebook — k-means corrects systematic binary bias

Target: ~1.18 bits per parameter, beating BiLLM on quality

---

The Method

Why Fisher > Hessian > Magnitude

Four Stages

FP32 Weights
    │
    ▼
Stage 1: Fisher-Weighted Channel Importance
    │
Stage 2: Mixed-Precision Core Allocation
    │  Top 5% channels → int4
    │  Bottom 95% channels → binary ±1
    ▼
Stage 3: Adaptive Blocksize Selection
    │  Per-layer blocksize {64, 128, 256, 512}
    ▼
Stage 4: Residual Codebook Clustering
    │  4 tiered codebooks × 64 centroids
    │  4-bit indices per block
    ▼
FABQ-RC Quantized Model
    │
    ▼
GGUF Export

Why Residual Codebook > Linear Approximation

BiLLM approximates residuals as a linear function of the weight value. FABQ-RC's k-means codebook is nonlinear and captures arbitrary residual patterns without assuming a functional form.

---

Quick Start

Download the Model

from huggingface_hub import snapshot_download

model_path = snapshot_download("toxzak/Qwen3.6-27B-FABQ-RC-GGUF")

Use with llama.cpp

# Example inference command
./llama-cli -m Qwen3.6-27B-FABQ-RC-Q4_K_M.gguf -n 256 -p "The future of 1-bit quantization is"

Evaluate

# Perplexity on WikiText-2
./llama-perplexity -m Qwen3.6-27B-FABQ-RC-Q4_K_M.gguf -f wikitext.txt

---

Model Details

---

Key Results

---

Files

fabq-rc/
├── README.md                              ← This file
├── FABQ_RC_SPEC.md                       ← Full method specification
├── FABQRC_PLAN.md                        ← Research plan
├── Main-FABQ-RC-Notebook.ipynb          ← Main quantization notebook
├── FABQ-RC-Dense-27B-Notebook.ipynb     ← Dense model experiments
└── plans/
    ├── CALIBRATION-ROBUSTNESS-PLAN.md  ← Calibration improvements
    ├── FABQ-VP-SPEC.md                  ← Variable precision extension
    ├── EBQ-SPEC.md                      ← Error-budget allocation
    └── UNIFIED-SPEC.md                   ← Combined architecture

---

Citation

FABQ-RC: Fisher-Adaptive Binary Quantization with Residual Codebooks
Zach Maronek, 2026

---

License

Apache 2.0 (see Hugging Face model page for details)

---

Built by Zach Maronek · April 2026