Google logo

Gemma 3n E4B

Multimodal
Google

Gemma 3n is a multimodal model designed for local hardware deployment, supporting image, text, audio, and video inputs. It includes a language decoder, audio encoder, and visual encoder and is available in two sizes: E2B and E4B. The model is optimized for efficient memory usage, allowing it to run on devices with limited GPU memory. Gemma is a family of lightweight, state-of-the-art open models from Google, built on the same research and technology used to create Gemini models. Gemma models are well-suited for various content understanding tasks, including question answering, summarization, and reasoning. Their relatively small size enables deployment in resource-constrained environments such as laptops, desktops, or private cloud infrastructure, democratizing access to state-of-the-art AI models and fostering innovation for everyone. Gemma 3n models are designed for efficient execution on resource-constrained devices. They can process multimodal inputs, working with text, image, video, and audio inputs, and generate text outputs, with open weights for instruction-tuned variants. These models were trained on data in over 140 spoken languages.

Key Specifications

Parameters
8.0B
Context
-
Release Date
June 26, 2025
Average Score
64.6%

Timeline

Key dates in the model's history
Announcement
June 26, 2025
Last Update
July 19, 2025
Today
March 25, 2026

Technical Specifications

Parameters
8.0B
Training Tokens
11.0T tokens
Knowledge Cutoff
June 1, 2024
Family
-
Capabilities
MultimodalZeroEval

Benchmark Results

Model performance metrics across various tests and benchmarks

General Knowledge

Tests on general knowledge and understanding
HellaSwag
10-shot AI: AI shows results at solving complex mathematical tasks with using n-shot examples. However n-shot at AI capability ability solve new tasks, since ensures its method evaluation uses 10-shot approach, in order to full AI, at this that examples substantially from test tasks, so that Examples and process solutions, but not specific test concepts. We prompt so: 1. for solutions mathematical tasks 2. 10 examples tasks and solutions from various fields 3. task This allows model diverse approaches to solving, maintaining at this and complexity tasksSelf-reported
78.6%
Winogrande
task (5-shot) - this provision example for language model, at which you five examples tasks, before than ask model execute her/its independently. This methodology "few-shot" (training), where number examples five. Demonstration five examples helps model better understand template or format output, especially for complex or tasks. Approach 5-shot gives model more context by comparison with 0-shot (without examples) or 1-shot (one example) that usually leads to more and results. When should use 5-shot approach: • For tasks, requiring specific format or structure • When necessary reasoning • For tasks, where is required specific or • When need to show solutions problems 5-shot approach is between provision too number examples (that can lead to to ) and too number (that can model and at main query)Self-reported
71.7%

Reasoning

Logical reasoning and analysis
BIG-Bench Hard
few-shotSelf-reported
52.9%
DROP
Token F1 score. 1-shot.Self-reported
60.8%

Other Tests

Specialized benchmarks
ARC-C
25-shotSelf-reported
61.6%
ARC-E
## 0-shot Model directly on test examples without additional instructions or training examplesSelf-reported
81.6%
BoolQ
0-shot (example) In model GPT, query "0-shot" means, that model not examples execution tasks. Model should rely exclusively on its preliminarily obtained knowledge and understanding instructions for formation answer. For example, in context mathematical tasks or tasks on reasoning, model simply is provided task without demonstration process solutions similar tasks. This k-shot, where model are provided examples (k examples) before that, how she/it solve task independently. performance model in conditions 0-shot is strict evaluation her/its basic abilitiesSelf-reported
81.6%
Natural Questions
5-shotSelf-reported
20.9%
PIQA
0-shot AI: is used without additional instructions, prompts or examples that, how execute task. For tasks 0-shot system with question. Example (in GPQA): Q: value expressions (8^67 + 27^41) mod 35? AI: In order to find value (8^67 + 27^41) mod 35, I I will use Advantages: - simple for settings - only instruction tasks - computational resources - basic abilities model Disadvantages: - Not gives model examples for training format or solutions - Can more iterations in solving - leads to to solvingSelf-reported
81.0%
Social IQa
shot capabilities model perform tasks without provision examples. you evaluate, how well model: • follow instructions • user • information at or For example, if model should answer only on queries about mathematics, and user question about whether model query. if user how solve equation, but not provides its, whether model equationSelf-reported
50.0%
TriviaQA
5-shotSelf-reported
70.2%

License & Metadata

License
proprietary
Announcement Date
June 26, 2025
Last Updated
July 19, 2025

Similar Models

All Models

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.