gemma-4-12B

gemma-4-12B

Google's Gemma 4 12B Unified model, an open model that reads text, images, audio and video and writes text, with a 256K context and no separate encoders. For developers running local AI.

Made by google on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: gemma-4-12B is the 12B Unified model in Google DeepMind's Gemma 4 family of open models. It takes in text, images, audio and video and writes text, and it has about 11.95 billion parameters.

The card calls it encoder-free: audio and vision are understood inside the one model, with no separate encoders.

Who it suits: Developers who want one local model for text, pictures and speech, and teams building assistants that run on consumer devices or workstations.

Industry | Who uses it | How they use it | Example
Software | App developers | Run a multimodal assistant locally | Ask a question about a photo
Research | AI researchers | Study an encoder-free multimodal model | Compare it with the other Gemma 4 sizes
Productivity | Tool builders | Build agents that call functions | Let the model call a calendar tool

Skip it if you need the model to draw images or speak aloud, since it writes text only.

Where to use it:

• Local setups with a consumer GPU or a workstation, which the card names as targets
• Hugging Face Transformers, which the library tag lists
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Reasoning and coding | Developers | Text answers, with thinking modes you can set
Image and audio questions | App builders | Text about a picture or a recording
Agents | Tool builders | Function calls, since the model supports calling functions natively

How to set it up:

• Python and pip to run the download commands
• About 23.9 GB of space for the BF16 weights, or about 6 GB at 4-bit by Hugging Face estimate
• A runtime that loads Transformers models, following the model card

1. Install the Hugging Face tools.
2. Download the model files.
3. Load the model in Transformers and follow the card's code.

pip install -U huggingface_hub
hf download google/gemma-4-12B

How to use it: Load the model and send it text, an image, audio or a mix, then read the text it writes. You can switch on its thinking mode and give it a system message to set the rules for the chat.

Property | 12B Unified
Parameters | 11.95B
Layers | 48
Sliding window | 1024 tokens
Context length | 256K tokens
Vocabulary | 262K
Inputs | Text, image, audio

Title: Size of the weights in GB
BF16 weights: 23.9 GB
4-bit estimate: 6 GB

Example: You give the model a photo of a whiteboard and ask it to turn the notes into a list. It reads the image itself and answers in text, with no separate vision tool.

For newcomers and experts: If you are new to multimodal models, start with text and one image before you try audio. Keep prompts short until you know how the thinking mode behaves.

If you know the field, layers alternate between local sliding-window attention and full global attention, and the last layer is always global. The global layers share Keys and Values and use proportional RoPE to save memory on long contexts.

Download the model files
Load the model in Transformers
Send text, an image or audio
Receive text, with optional thinking

Good to know:

• The card says audio is supported natively on E2B, E4B and 12B
• The model supports the system role for more structured chats
• Hugging Face lists 51 quantized and 96 fine-tuned versions made by other people
• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the files you download and the text, images and audio you give it
• Not for: making images or audio, since the output is text

On Hugging Face, 13 November 2026

• Task: any to any
• Library: transformers
• Parameters: 12B
• Weights: about 23.9 GB in BF16; about 6 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 51,228
• Likes: 861
• License: apache-2.0
• Quantized versions: 51
• Fine-tunes: 97
• Tags: transformers, safetensors, gemma4_unified, image-text-to-text, any-to-any

Links

• Model: https://huggingface.co/google/gemma-4-12B
• Files: https://huggingface.co/google/gemma-4-12B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download google/gemma-4-12B

Then follow the usage in its model card. Nothing in this kit runs by itself.