gemma-4-31B

gemma-4-31B

Reads text and images and writes text, for developers and researchers who want Google DeepMind's largest dense open-weights Gemma 4 model for coding, reasoning and image questions.

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 31B is the largest dense size in the open Gemma 4 family from Google DeepMind. It reads text and images and writes text. The repository holds about 31.3B parameters as Safetensors files for the transformers library, and it has no audio input.

Who it suits: It suits developers and researchers who want an open-weights model for text generation, coding, reasoning and image understanding, run on a consumer GPU or a workstation.

Title: Who uses this model
Industry | Who uses it | How they use it | Example
Software | Developers | Add coding help and reasoning to their own apps | Ask it to review a function
Research | Machine learning researchers | Study a dense open-weights model | Compare it with the smaller sizes in the family
Product teams | Agent builders | Build assistants that call functions | Give the model a tool and let it decide when to use it

Anyone who needs audio input can skip this size, since the card lists audio only for the E2B, E4B and 12B models.

Where to use it:

• Python scripts that use the transformers library
• A consumer GPU or a workstation
• Windows, macOS or Linux with Python, for downloading the files

Title: Jobs it does
Use case | Who | What they get back
Text generation | Writers and developers | Written replies to a prompt
Coding help | Developers | Code and explanations in text
Image questions | Anyone with pictures to check | Text answers about an image you supply
Reasoning | Researchers | Step-by-step answers, with thinking modes you can configure

How to set it up: You need Python and plenty of disk space. The BF16 files weigh about 62.5 GB, and a 4-bit copy is estimated at about 15.6 GB.

• Python with pip
• The huggingface_hub package
• Free disk space for the weights

1. Install huggingface_hub.
2. Run the download command to save the model files on your computer.
3. Load the files with the transformers library in your own script.

pip install -U huggingface_hub
hf download google/gemma-4-31B

How to use it: Load the downloaded files with transformers and give the model text, or an image together with text. The system role lets you set ground rules for a conversation. The card says thinking modes are configurable and points to its own pages for the exact settings.

Example: You hand the model a screenshot of an error message and ask what it means in plain words. It answers in text. A system message can tell it to keep replies short and formal.

For newcomers and experts: Think of it as a model that looks at a picture and a question, then writes an answer. A first small test is one image and one short question. The family also has smaller sizes if this one is too big for your computer.

The dense design means all parameters work on every token. The card gives 60 layers, a sliding window of 1024 tokens, a 256K-token context length and a vocabulary of 262K. Attention alternates between local windows and full global layers, with the last layer always global. Global layers share keys and values and use Proportional RoPE to save memory. The vision encoder has roughly 550M parameters.

Text or image goes in
Vision encoder reads the image
Hybrid attention layers
Text comes out

Good to know: The family reaches a context window of 256K tokens and supports more than 140 languages. Other people have published 44 quantized and 103 fine-tuned versions of this model. The release has pre-trained and instruction-tuned variants, so check the repository name to see which one you get.

Title: Estimated weight size
BF16: 62.5 GB
4-bit: 15.6 GB

• License: Apache 2.0.
• Cost to run: none named.
• Data it reaches: the files come from Hugging Face, and the model then runs on your own hardware.
• Not for: audio input, which the card lists only for E2B, E4B and 12B.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 31.3B
• Weights: about 62.5 GB in BF16; about 15.6 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 44,151
• Likes: 596
• License: apache-2.0
• Quantized versions: 46
• Fine-tunes: 103
• Tags: transformers, safetensors, gemma4, image-text-to-text

Links

• Model: https://huggingface.co/google/gemma-4-31B
• Files: https://huggingface.co/google/gemma-4-31B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download google/gemma-4-31B

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