Modelpublic

positron-ai/google_gemma-3-27b-it-ingest-best-gptq

sourceHugging Faceotherupdated 1mo agoView on Hugging Face

0likes555downloads

Model Card

Positron AI Quantized Build

This repository contains a Positron AI quantized build of google/gemma-3-27b-it for inference.

Use this artifact when you need a GPTQ 4-bit build of google/gemma-3-27b-it built by Positron AI.

For general-purpose GPU inference, compare against the original model and other quantized formats before deployment.

Field	Value
Base model	google/gemma-3-27b-it
Published artifact	positron-ai/google_gemma-3-27b-it-ingest-best-gptq
Quantization method	GPTQ
Quantization format	gptq
Source precision	n/a
Target runtime	n/a
Hardware target	n/a
Release date	2026-08-25
License	other

Field	Value
Weight precision	4-bit
Activation precision	not quantized
Bits	4
Group size	64
Symmetric quantization	true
Activation ordering / desc_act	false
Damp percent	0.05
Calibration dataset	Mixed-domain calibration set
Calibration samples	128
Calibration sequence length	4096
MoE experts per token	n/a
Quantization toolchain	GPTQModel 7.2.0, transformers 5.11.0, torch 2.9.1, CUDA 12.8

Metric	Result	Reference	Notes
Mean KL-divergence	n/a	n/a	Not measured for this release
P95 KL-divergence	n/a	n/a	Not measured for this release
Top-1 agreement	n/a	n/a	Not measured for this release
Perplexity / NLL delta	n/a	n/a	Not measured for this release
MMLU mean	pending	n/a	Evaluation pending

KL-divergence has not been measured for this release; the metrics above will be populated when validation completes.

This artifact was produced by Positron AI from google/gemma-3-27b-it. The original model license and usage restrictions continue to apply.