Gemma 3n E2B
MultimodalGemma 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 RAM. 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, images, video, and audio, 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
58.6%
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 method more high performance by comparison with type few-shot. In difference from methods, which provide model only in advance examples for understanding tasks, 10-shot allows model first 10 examples, and then use these examples how at solving new problems. Process 10-shot allows model conduct more analysis examples and connection between them and new task, that more reasoning. This method especially efficient for complex tasks, requiring logical thinking, such how mathematical tasks, puzzles or reasoning. 10-shot includes provision model set from 10 various examples with detailed solutions, which she/it can use in capacity at execution new tasks. This allows model templates and strategies solutions, which can apply to • Self-reported
Winogrande
5-shot AI: GPT-4 and relying on on 5 examples. Solution problems for 5 steps: 1. Example 1 → training solving specific tasks 2. Example 2 → understanding new 3. Example 3 → processing cases 4. Example 4 → application 5. Example 5 → and verification conclusions Efficiency: - accuracy on 22-30% in complex tasks by comparison with 1-shot - errors in mathematical reasoning - More understanding context and Application: • for tasks, requiring reasoning • at training complex • Outperforms other approaches in tasks with high Limitations: • tokens • examples • Can be for simple tasks • Self-reported
Reasoning
Logical reasoning and analysis
BIG-Bench Hard
few-shot • Self-reported
DROP
Token F1 score. 1-shot. • Self-reported
Other Tests
Specialized benchmarks
ARC-C
25-shot • Self-reported
ARC-E
0-shot AI: (I "I") For solutions this tasks me need to determine, which can on 5×5 so, in order to not with : - : 1 in : number by or : number by or : number by : course "" (2 by one then 1 ) - : on 1 by For number I should effectively use considering field each : 1. with most () 2. Then with () 3. that not with and so how their field most Then which have and if I would about 8-10 but analysis for exact answer • Self-reported
BoolQ
Models with 0-shot trials on standard benchmarks without additional We simply model task from test and how she/it handles. For each test is used standard instruction with tasks. For example, model receives task "Solve equation 2x + 5 = 13" and should answer, that x = 4. Important note, that all our data about models in 0-shot mode and thoroughly our in order to in their accuracy • Self-reported
Natural Questions
5-shot training, also how "training in context" (in-context learning, ICL), allows models (LLM) new tasks from several examples, in prompt, without additional training. In this work we question about that, how LLM computation for solutions tasks. We efficiency different instructions by solutions for mathematical and tasks, from query final answer to lead to reasoning. We we consider five instructions: 1) standard, 2) justification answer, 3) reasoning, 4) only answer, and 5) explanation steps after obtaining correct answer. We we evaluate these instructions on tasks from five various mathematical and and we verify three model: GPT-3.5 Turbo, GPT-4, and Claude 2. In dependency from complexity tasks reasoning can performance on 6-53% by comparison with only answer. results show, that reasoning can lead to to on tasks, but significantly performance on complex tasks. solutions, such how time answer, can for when reasoning will We also that performance model can be if not use reasoning on tasks and use their on complex tasks • Self-reported
PIQA
0-shot prompt - this when model uses its in order to directly answer on question, without examples similar questions. This reflects performance model without any-or additional instructions about that, how answer on specific questions. 0-shot approach especially for new or tasks, where model should rely only on its basic knowledge and abilities • Self-reported
Social IQa
Method 0-shot (or "") - this way testing model artificial intelligence, at which model with new task without any-or preliminary examples or instructions about that, how her/its solve. Model should exclusively on knowledge, in time preliminary training. In 0-shot testing, model is provided only instruction and task, which need to execute. For example, model can obtain question or problem without additional context, examples solutions or demonstrations. This method especially important for evaluation abilities model and her/its abilities. 0-shot testing allows understand, how well well model can apply its knowledge to new or tasks, that is intelligence. Although other methods, such how few-shot (with several examples) can improve performance on specific tasks, 0-shot testing often is considered more abilities model, since it capability that, that model simply templates from examples • Self-reported
TriviaQA
5-shot Method 5-shot based on that results LLM can be by means of provision several examples execution tasks before that, how model solve her/its independently. Approach: 1. For each tasks from 5 similar tasks that indeed type and level complexity. 2. Models show 5 examples together with their detailed solutions. 3. after this model solve task. Advantages method: • Allows model and templates solutions tasks specific type • Not requires model • improves results on complex mathematical and logical tasks • Can be for specific knowledge Limitations: • examples • that can be for models with context • Examples should be indeed this can model with : Especially efficient at in with Chain-of-Thought prompting, where 5 examples demonstrate not only answers, but and full course reasoning • Self-reported
License & Metadata
License
proprietary
Announcement Date
June 26, 2025
Last Updated
July 19, 2025
Similar Models
All ModelsGemma 3n E2B Instructed
MM8.0B
Best score:0.7 (HumanEval)
Released:Jun 2025
Gemma 3 4B
MM4.0B
Best score:0.7 (HumanEval)
Released:Mar 2025
Price:$0.02/1M tokens
Gemma 3n E4B Instructed LiteRT Preview
MM1.9B
Best score:0.8 (HumanEval)
Released:May 2025
Gemma 3n E2B Instructed LiteRT (Preview)
MM1.9B
Best score:0.7 (HumanEval)
Released:May 2025
Gemma 3n E4B Instructed
MM8.0B
Best score:0.8 (HumanEval)
Released:Jun 2025
Price:$20.00/1M tokens
Gemini 1.5 Flash 8B
MM8.0B
Best score:0.4 (GPQA)
Released:Mar 2024
Price:$0.07/1M tokens
MedGemma 4B IT
MM4.3B
Released:May 2025
Gemma 3n E4B
MM8.0B
Best score:0.6 (ARC)
Released:Jun 2025
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.