Robots Atlas>ROBOTS ATLAS
Gemma-3-4B-it
AI Modelsโ€บGemma

Gemma-3-4B-it

3 ยท 4B ยท Instructย ยทย Family: Gemma
An open, instruction-tuned 4B multimodal model from Google DeepMind. Text+image input, 128K context, 140+ languages. Runs on a laptop. Gemma license.
โœ“ Activeโœ“ Public accessโš– Open weightsMultimodalLLM๐Ÿ“ Gemma
Context window
128K
tokens
Parameters
4B
parameters
Max output
8,192
tokens
Release date
12 March 2025
๐ŸขGoogle DeepMindProducer๐ŸขGoogleProducer
Access:DownloadHostedDeployment:๐Ÿ’ป Localโ˜ Cloud๐Ÿ“ฑ On-device

Overview

Gemma-3-4B-it is an instruction-tuned, multimodal vision-language model with 4B parameters, developed by Google DeepMind. It accepts text and images (normalized to 896ร—896 and encoded as 256 tokens each) and generates text responses of up to 8,192 tokens.

The model supports a 128K-token context window and over 140 languages. It was trained on 4 trillion tokens (web documents, code, mathematics, images) with rigorous sensitive-data filtering, using JAX and ML Pathways on TPU hardware (TPUv4p/v5p/v5e).

Gemma-3-4B-it excels at question answering, summarization, reasoning, and image understanding; sample results include MMLU 59.6% (5-shot) and HumanEval 36.0% (0-shot). It is designed for deployment in resource-constrained environments (laptops, personal cloud). Released under the Gemma license (requires acceptance of Google's terms).

Classification
MultimodalLLM
Family: Gemma
Access & deployment
DownloadHosted
LocalCloudOn-device
Weights: Open weights
Key parameters
๐Ÿ“ Context: 128K
๐Ÿงฉ Parameters: 4B
โœ“ Toolsย ยทย โœ“ Fine-tuning
๐Ÿ“ฅ Input: text, image

Technical specification

Context window
128K
tokens
Parameters
4B
parameters
Max output tokens
8,192
tokens per response
License
Gemma
Hardware requirements
Designed for resource-constrained devices (laptop, personal cloud). Image input 896ร—896 โ†’ 256 tokens.
Features:โœ“ Tool useโœ“ Fine-tuning
Modalities
โฌ‡ Input
textimage
โฌ† Output
textcode

Capabilities and applications

Native model capabilities
Multimodal understanding
Category: multimodal
Image understanding
Analysing and interpreting the content of images.
Category: vision
Reasoning
The model's ability to reason logically and solve complex problems.
Category: reasoning
Multilingual
Competence in many natural languages (from a few to over a hundred): understanding, generation, translation, and code-switching within a single conversation. Frontier models support a wide range of languages with comparable quality.
Category: language
Instruction following
Precisely following instructions contained in the prompt: response format, length, style, constraints (e.g. 'reply in six words'). GPT-5.1 significantly improved this capability compared to GPT-5.
Category: language
Long context
Support for large context windows โ€” tens to hundreds of thousands (or millions) of input tokens. Enables analysis of entire codebases, long documents, and many parallel conversations without losing earlier information. GPT-5.1 supports 400,000 tokens.
Category: language
Coding
Generating, analysing and modifying code in many programming languages. Covers writing functions, debugging, refactoring, code review, and creating tests. Measured by benchmarks such as HumanEval and SWE-bench.
Category: coding

Benchmark results

2 benchmarks
MMLU
accuracy ยท 5-shot
59.6%
๐Ÿ“„ technical_report
HumanEval
pass@1 ยท 0-shot
36.0%
๐Ÿ“„ technical_report

Technical architecture

Core Architecture
TRTransformer
Training Techniques
ITInstruction TuningRLRLHF

Sources and related pages