gemma-4-12B-it

gemma-4-12B-it

Takes text, images and audio and writes text answers, in one 12B instruction-tuned model from Google DeepMind with a 256K-token context and configurable thinking. For developers running open models locally.

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-it is the instruction-tuned 12B Unified model of Google DeepMind's Gemma 4 open family. It understands audio and images natively, without separate encoders, and writes text. The context window reaches 256K tokens, and the family covers more than 140 languages.

Who it suits: It suits developers who want a capable open model on a consumer GPU or workstation for reasoning, coding and multimodal work.

Industry | Who uses it | How they use it | Example
Software development | Developers | Use it for coding help and agent workflows with function calling | Asking it to fix a failing test
Accessibility and media | App builders | Process spoken audio and images in one model | Describing a photo and answering a question about it
Research | Researchers | Fine-tune or evaluate an open model with a long context | Testing long-document question answering
Multilingual products | Product teams | Serve users in many languages from one local model | Answering a support question in another language

You can skip it if you need a phone-sized model, since the family has smaller E2B and E4B sizes for that.

Where to use it:

• Python with the transformers library
• Consumer GPUs and workstations, which the card names as the target for this size
• 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 configure
Audio and image questions | App builders | Text answers about what the model hears or sees
Agent workflows | Agent developers | Native function calls and system-role instructions

How to set it up: You need Python and a machine that can hold the weights, about 23.9 GB in BF16 or roughly 6 GB at 4-bit as an estimate.

• Python on Windows, macOS or Linux
• The huggingface_hub package
• The transformers library to run the model
• A consumer GPU or workstation with enough memory

1. Install huggingface_hub.
2. Download the model files to your computer.
3. Load the model with transformers and follow the run steps on the model page.

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

How to use it: The visible card gives the download step and points to its own run section. You load the instruction-tuned checkpoint, write a system message to set behavior, and send text, images or audio.

hf download google/gemma-4-12B-it

Example: You send the model a photo and ask what is in it. The reply comes back as text, and you can also pick a thinking mode if you want it to reason before it answers.

For newcomers and experts: If you are new to open models, this is a mid-sized one that handles pictures and sound as well as text. A first small try is to download it and ask one short question in plain text.

For people who know the field, the model has 48 layers and a hybrid attention design that mixes local sliding windows of 1024 tokens with global layers, ending on a global one. Being encoder-free is a trade-off in size and deployment, since audio and vision pass through the model itself.

Text, image or audio goes in
The unified model handles all inputs together
It may think first if a thinking mode is on
You get a text answer

Good to know: The Gemma 4 line offers five models, from the phone-class E2B and E4B up to the 31B, and this 12B is one of the three that hear audio natively.

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the text, images and audio you send, plus the downloaded model files
• Not for: audio generation, since the model writes text output

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: 1,738,905
• Likes: 1,735
• License: apache-2.0
• Based on: google/gemma-4-12B
• Quantized versions: 403
• Fine-tunes: 306
• Tags: transformers, safetensors, gemma4_unified, image-text-to-text, any-to-any

Links

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

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