gemma-4-31B-it

gemma-4-31B-it

An instruction-tuned 31B model from Google DeepMind that reads text and images and writes text, with a 256K context and configurable thinking. For developers who run an open model on a workstation or server.

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-it is the instruction-tuned 31B dense model of Gemma 4, an open model family from Google DeepMind. It takes text and images and writes text. It has a context window of 256K tokens and supports more than 140 languages. Its weights are open under Apache 2.0.

Who it suits: It suits developers, researchers and teams who want an open model for chat, coding, reasoning and agents, hosted on their own GPU or server.

Industry | Who uses it | How they use it | Example
Software | Developers | Add a multimodal assistant to an app | A helper that answers questions about screenshots
Global business | Support and content teams | Work in many languages with one model | Reply to customers in their own language
Agent building | Agent developers | Use function calling to drive tools | An agent that looks up orders through an API
AI research | Researchers | Study an open reasoning model | Compare thinking modes on the same questions

If you need audio input or a model for a phone, you can skip it, because audio sits only in the smaller sizes and the 31B has none.

Where to use it:

• Python with the transformers library
• Consumer GPUs, workstations and servers, the range the card gives for the larger sizes
• Windows, macOS or Linux, with enough memory for a 31B model
• Agent setups that need native function calling

Use case | Who | What they get back
Chat and reasoning | Developers | Text answers, with a thinking mode you can configure
Image questions | Product teams | Text answers about an image of any shape or size
Tool-driven agents | Agent builders | Function calls the model chooses to make

How to set it up: You need these first:

• Python and pip
• A GPU setup with room for about 62.5 GB of BF16 weights, or about 15.6 GB at a 4-bit estimate

1. Install the Hugging Face hub tool.
2. Download the model files.
3. Load the model with transformers as the model card describes.

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

How to use it: Chat with the model through its chat template, using the system role for instructions. Pick a thinking mode when you want step-by-step reasoning, and send an image along with your text when you have one.

Tool | What it does
System role | Native support for a system message, for more controllable conversations
Thinking modes | Configurable reasoning before the final answer
Function calling | Native support for calling tools, which agents rely on
Image input | Takes images with a variable aspect ratio and resolution

Example: You send a photo of a receipt and ask for the total and the shop name. The model reads the image and answers in text. With a system message, you can also fix the answer format.

For newcomers and experts: This is a large general model that understands pictures as well as words. A good first step is to read the model card, then try one text question and one image question.

The model mixes local sliding window attention, 1024 tokens wide, with full global attention, and the last layer is always global. Global layers share keys and values and use proportional RoPE. It has 60 layers and a vision encoder of about 550M parameters.

You send text and optionally an image
The model reads them with a system role
It may think before it answers
It returns text or a function call

Good to know: The Gemma 4 family has five sizes, and this one is the largest dense model.

Title: Total parameters of the dense Gemma 4 models in billions
E2B (effective): 2.3 B
E4B (effective): 4.5 B
12B Unified: 11.95 B
31B Dense: 30.7 B

• Sizes: five in all, with this 31B at the top and E2B, E4B, 12B and 26B A4B below it, where small models have a 128K context and medium ones 256K
• Modalities: this 31B model takes text and images, with no audio
• Other versions: 358 quantized and 369 fine-tuned versions made by other people are listed
• License: Apache 2.0
• Cost to run: none named, you supply the GPU
• Data it reaches: the prompts and images you give it

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: 6,692,367
• Likes: 4,348
• License: apache-2.0
• Based on: google/gemma-4-31B
• Quantized versions: 363
• Fine-tunes: 372
• Tags: transformers, safetensors, gemma4, image-text-to-text, conversational

Links

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

Get the model

Download its files with Hugging Face's command line:

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

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