Modelpublic

rshoemake/taq-qwen3-coder-next-outlier-4bit

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes15downloads

Model Card

TAQ Qwen3-Coder-Next — outlier-preserving 4-bit research scaffold

This repository contains the completed Tail-Aware Quantization (TAQ) outlier_4bit research checkpoint derived from Qwen/Qwen3-Coder-Next.

Important storage/deployment note

The tensors in this repository are the dequantized BF16/FP16 evaluation scaffold produced by the research pipeline. They are loadable with Transformers, but they occupy approximately the original checkpoint size on disk. This repository is not yet a packed 4-bit deployment format and does not claim a physical VRAM or disk-size reduction.

The quantization accounting measured during construction is:

—Total parameters: 79,674,391,296
—Quantized parameters: 77,787,561,984 (97.63%)
—Untouched parameters: 1,886,829,312
—Exact whole-model logical storage: 4.4452874061 bpw
—Logical packed bytes: 44,271,946,028
—Logical compression versus fp16: 3.5993173305×
—Physical dequantized scaffold bytes: 159,357,958,880
—Group size: 128
—Outlier fraction: 0.005
—K-means iterations: 20

The packed-kernel representation is separate future work; do not interpret the logical bpw above as this repository's physical file size or runtime footprint.

Evaluation state

—Full-precision reference PPL (WikiText-2, n=40): 12.376617415789612
—Quantized PPL, HumanEval+, and MBPP+ are pending matched evaluation.

Base model

See Qwen/Qwen3-Coder-Next for architecture, usage, license, and original model documentation.