gemma-4-26B-A4B-it-GEM16

danmoreng/gemma-4-26B-A4B-it-GEM16

gemma-4-26B-A4B-it-GEM16 is a image-text-to-text model from danmoreng. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for gem16. The card lists the license as apache-2.0.

This repository contains the GEM16-native components for the Gemma 4 26B A4B Compact Vision profile on NVIDIA Blackwell SM120. The payload filenames state the real model and storage/quantization format; there is no an…

Downloads · 30 days

0% of all-time downloads

All-time downloads

772

Public

Repo size

40 GB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.gem1627.5 GB · 99%

At a glance

Task: Image-Text-to-Text
Library: gem16
License: apache-2.0
Access: Public
Created: Aug 27, 2026
Updated: Sep 4, 2026
SHA: 5a7a0225

Try a prompt

Base models

google/gemma-4-26B-A4B-it-qat-q4_0-unquantized

Model card

From the Hugging Face model README

Gemma 4 26B A4B IT — GEM16

This repository contains the GEM16-native components for the Gemma 4 26B A4B Compact Vision profile on NVIDIA Blackwell SM120. The payload filenames state the real model and storage/quantization format; there is no anonymous model.gem16 payload in this revision.

Role	Artifact	Quantization / storage	Product status
`text-target`	`components/gemma-4-26b-a4b-it-trellis35-w4a8/gemma-4-26b-a4b-it-trellis35-w4a8.gem16`	Trellis35 W4A8	public component
`vision-encoder`	`components/gemma-4-26b-a4b-it-vision-fp8-e4m3fn/gemma-4-26b-a4b-it-vision-fp8-e4m3fn.gem16`	FP8 E4M3FN rowwise	public component
`fixed-d2-assistant`	`components/gemma-4-26b-a4b-it-assistant-hybrid-nvfp4-fp8-bf16/gemma-4-26b-a4b-it-assistant-hybrid-nvfp4-fp8-bf16.safetensors`	hybrid NVFP4 / FP8 / BF16	public component
`internal-text-target`	`internal/gemma-4-26b-a4b-it-target-hybrid-nvfp4-fp8-bf16/gemma-4-26b-a4b-it-target-hybrid-nvfp4-fp8-bf16.gem16`	hybrid NVFP4 / FP8 / BF16	internal regression/rollback

The public profile combines the Trellis35 W4A8 text Target with the FP8 E4M3FN Vision component. The fixed-D2 Assistant is optional. The larger hybrid NVFP4 Target remains available under internal/ for regression and rollback, but it is not a normal Studio product choice.

Trellis35 is EXL3-derived offline packing in GEM16's native direct-load format; its routed experts use mixed K3/K4 storage at approximately 3.5 bits per weight. Vision uses FP8 E4M3FN linear weights with BF16 row scales and BF16 support tensors—it is not Q8. The Assistant and internal Target are hybrid NVFP4/FP8/BF16 artifacts.

profiles.json is the machine-readable composition contract. Each component has a component.json with its exact artifact identity and quantization. The compilation.json and source.lock.json files preserve compiler and immutable source provenance.

These are compiled GEM16 artifacts, not Transformers, GGUF, or generic Safetensors checkpoints. Runtime model hashing is deliberately disabled; integrity is established during download/install and recorded in the locks.

Engine and usage documentation: https://github.com/Danmoreng/gem16

Task: Image-Text-to-Text
Library: gem16
License: apache-2.0
Created: Aug 27, 2026
Updated: Sep 4, 2026