Model Information for Gemma 3

Model Page

Gemma

Resources and Technical Documentation

• Gemma 3 Technical Report
• Responsible Generative AI Toolkit
• Gemma on Kaggle
• Gemma on Vertex Model Garden
• Terms of Use: Terms

Authors: Google DeepMind

---

Model Information

Description

Gemma is a family of lightweight, state-of-the-art open models from Google, built from the same research and technology used to create the Gemini models.

Gemma 3 models are multimodal, handling text and image input and generating text output, with open weights for both pre-trained variants and instruction-tuned variants.

• Context Window: 128K tokens (4B, 12B, 27B), 32K tokens (1B)
• Multilingual Support: Over 140 languages
• Deployability: Designed for laptops, desktops, and private cloud infrastructure

Applications:
Question answering, summarization, reasoning, image understanding, and more.

---

Inputs and Outputs

Input:

• Text string: questions, prompts, documents
• Images: normalized to 896x896 resolution, encoded to 256 tokens each
• Total input context: 128K or 32K tokens (depending on model size)

Output:

• Generated text
• Total output context: 8192 tokens

---

Usage

Installation

1pip install -U transformers

Running with the Pipeline API

1from transformers import pipeline
2import torch
3
4pipe = pipeline(
5    "image-text-to-text",
6    model="google/gemma-3-27b-it",
7    device="cuda",
8    torch_dtype=torch.bfloat16
9)
10
11messages = [
12    {
13        "role": "system",
14        "content": [{"type": "text", "text": "You are a helpful assistant."}]
15    },
16    {
17        "role": "user",
18        "content": [
19            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
20            {"type": "text", "text": "What animal is on the candy?"}
21        ]
22    }
23]
24
25output = pipe(text=messages, max_new_tokens=200)
26print(output[0]["generated_text"][-1]["content"])

Example Output:

Based on the image, the animal on the candy is a turtle.

---

Running on Single/Multi-GPU

1# pip install accelerate
2from transformers import AutoProcessor, Gemma3ForConditionalGeneration
3from PIL import Image
4import requests
5import torch
6
7model_id = "google/gemma-3-27b-it"
8
9model = Gemma3ForConditionalGeneration.from_pretrained(
10    model_id, device_map="auto"
11).eval()
12
13processor = AutoProcessor.from_pretrained(model_id)
14
15messages = [
16    {
17        "role": "system",
18        "content": [{"type": "text", "text": "You are a helpful assistant."}]
19    },
20    {
21        "role": "user",
22        "content": [
23            {"type": "image", "image": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"},
24            {"type": "text", "text": "Describe this image in detail."}
25        ]
26    }
27]
28
29inputs = processor.apply_chat_template(
30    messages, add_generation_prompt=True, tokenize=True,
31    return_dict=True, return_tensors="pt"
32).to(model.device, dtype=torch.bfloat16)
33
34input_len = inputs["input_ids"].shape[-1]
35
36with torch.inference_mode():
37    generation = model.generate(**inputs, max_new_tokens=100, do_sample=False)
38    generation = generation[0][input_len:]
39
40decoded = processor.decode(generation, skip_special_tokens=True)
41print(decoded)

---

Citation

1@article{gemma_2025,
2    title={Gemma 3},
3    url={https://goo.gle/Gemma3Report},
4    publisher={Kaggle},
5    author={Gemma Team},
6    year={2025}
7}

---

Model Data

Training Dataset

• Text: Diverse web documents in over 140 languages
• Code: Exposure to programming languages
• Mathematics: Logical reasoning and symbolic representation
• Images: Broad image data for vision tasks

Model Size	Tokens Used for Training
27B	14 trillion
12B	12 trillion
4B	4 trillion
1B	2 trillion

---

Data Preprocessing

• CSAM Filtering: To exclude harmful content
• Sensitive Data Filtering: Removal of certain personal information
• Content Quality Filtering: Following Google policies

---

Implementation Information

Hardware

• Trained on TPUv4p, TPUv5p, TPUv5e
• Advantages:
  • High performance for matrix operations
  • Large high-bandwidth memory
  • Scalability via TPU Pods
  • Cost-effective training at large scales

Software

• Frameworks: JAX + ML Pathways
• Benefits:
  • Single-controller programming model
  • Efficient orchestration of large-scale training

---

Evaluation

Reasoning and Factuality

Benchmark	Metric	1B	4B	12B	27B
HellaSwag	10-shot	62.3	77.2	84.2	85.6
BoolQ	0-shot	63.2	72.3	78.8	82.4
PIQA	0-shot	73.8	79.6	81.8	83.3
SocialIQA	0-shot	48.9	51.9	53.4	54.9
TriviaQA	5-shot	39.8	65.8	78.2	85.5
Natural Questions	5-shot	9.48	20.0	31.4	36.1
ARC-c	25-shot	38.4	56.2	68.9	70.6
ARC-e	0-shot	73.0	82.4	88.3	89.0
WinoGrande	5-shot	58.2	64.7	74.3	78.8
BIG-Bench Hard	few-shot	28.4	50.9	72.6	77.7
DROP	1-shot	42.4	60.1	72.2	77.2

STEM and Code

Benchmark	Metric	4B	12B	27B
MMLU	5-shot	59.6	74.5	78.6
AGIEval	3-5-shot	42.1	57.4	66.2
GSM8K	8-shot	38.4	71.0	82.6
HumanEval	0-shot	36.0	45.7	48.8

Multilingual

Benchmark	1B	4B	12B	27B
MGSM	2.04	34.7	64.3	74.3
Global-MMLU-Lite	24.9	57.0	69.4	75.7
FloRes	29.5	39.2	46.0	48.8

Multimodal

Benchmark	4B	12B	27B
COCOcap	102	111	116
DocVQA (val)	72.8	82.3	85.6
TextVQA (val)	58.9	66.5	68.6
RealWorldQA	45.5	52.2	53.9

---

Ethics and Safety

Evaluation Approach

• Structured evaluation and internal red-teaming
• Focus areas:
  • Child Safety
  • Content Safety
  • Representational Harms

Results

• Major improvements over earlier Gemma models
• Minimal policy violations observed
• All testing without additional safety filters

---

Usage and Limitations

Intended Usage

• Content creation
• Chatbots and conversational AI
• Text summarization
• Image data extraction
• Research and education

Limitations

• Potential biases from training data
• Challenges with open-ended or ambiguous tasks
• Occasional factual inaccuracies
• Limited common sense reasoning
• English-centric safety evaluation

---