Google Gemma 4 LiteRT-LM

Gemma 4 E2B on Android Offline.

Multimodal text, vision, and audio understanding in a single on-device model

Gemma 4 E2B is Google's multimodal foundation model, packaged for Google AI Edge (LiteRT-LM). In OfflineGPT, it runs natively on your Android device to process text, image attachments, and spoken audio in a single context without sending data to the cloud.

Size: ~2.58 GB Format: LiteRT-LM (.litertlm) Minimum RAM: 8 GB

Specifications

Technical Specifications

Verified technical parameters for the official .litertlm artifact distributed in OfflineGPT.

Model Name	Gemma 4 E2B (Gemma-4-E2B-it)
Developer / Provider	Google
Inference Engine	LiteRT-LM
Artifact Format	`LiteRT-LM (.litertlm)`
Download Size	~2.58 GB (2,583,085,056 bytes)
RAM Requirement	Min: 8 GB RAM
Hardware Acceleration	OpenCL GPU & Snapdragon NPU
Supported Modalities	TextVision (Images)Audio Input
Context Window	Adaptive on-device cap (up to 32,768 tokens)
File Name in App	`gemma-4-E2B-it.litertlm`

Hardware

Hardware & RAM Requirements

Device memory thresholds and acceleration engines required for smooth on-device execution.

RAM Requirement: 8 GB Minimum

8 GB RAM is the minimum requirement for Gemma 4 E2B due to the combined memory footprint of the weights (~2.58 GB), KV cache, and multimodal visual/audio buffers. Phones with 6 GB or less cannot allocate sufficient contiguous memory and may crash. OpenCL GPU acceleration is supported on Adreno and Mali GPUs with standard vendor drivers.

Native LiteRT-LM Engine

Executes via Google LiteRT-LM C++ runtime with hardware acceleration options for OpenCL GPU and Qualcomm Snapdragon NPU.

Single-Bundle Artifact

Packaged as a verified 2.58 GB .litertlm bundle containing the text model, vision encoder, audio projector, and SentencePiece tokenizer sidecar.

Multimodal Context

Accepts mixed inputs in one turn: attach a photo of a circuit or chart, speak a voice question, and receive structured text analysis.

Adaptive KV Memory Management

Dynamically sizes KV cache allocations to prevent Android low-memory killer (LMK) terminations during multi-turn chats.

Architecture

How Gemma 4 Runs on Android

On-device LiteRT-LM runtime, tokenizer integration, and multi-modal pipeline.

Gemma 4 E2B is Google's multimodal on-device foundation model, specifically quantized and packaged for Google AI Edge (LiteRT-LM). In OfflineGPT, it runs through the native LiteRT-LM runtime on Android, allowing simultaneous understanding of text prompts, camera or gallery photos, and audio recordings.

Unlike conventional multi-engine setups that pass image crops through an external vision projector into a separate text LLM, Gemma 4 processes image tokens and audio tokens directly in its unified architecture. This reduces translation overhead and preserves contextual nuances across modalities.

In OfflineGPT, Gemma 4 E2B also integrates with the on-device web search tool. When enabled, the model can query search results to answer questions requiring current facts, while still running inference entirely on the smartphone.

Use Cases

Practical On-Device Use Cases

Real tasks where multimodal on-device understanding provides immediate utility without internet access.

Technical Document & Diagram Analysis

Inspect wiring diagrams, mathematical plots, architectural blueprints, or error screenshots directly from the device camera or gallery.

Voice-Prompted Visual Inspection

Record audio questions while pointing the camera at physical equipment, receipts, or plant specimens without typing.

Offline Travel & Multilingual Sign Translation

Capture photos of street signs, foreign menus, or informational brochures in airplane mode and ask for detailed contextual translations.

Web-Assisted Research on Device

Toggle on the on-device search tool to let Gemma 4 E2B ground its answers with live web excerpts when Wi-Fi or mobile data is available.

Tradeoffs

Known Limitations & Tradeoffs

Realistic operational constraints for running a foundation model on mobile hardware.

Strict 8 GB RAM requirement: Devices with 4 GB or 6 GB of RAM will experience system memory pressure or app terminations.
Audio turns require substantial working memory: Long audio snippets consume additional input tokens and buffer space.
Initial download size: The .litertlm bundle is ~2.58 GB and should be downloaded over a reliable Wi-Fi connection.
Thermal profile on older chipsets: Sustained multimodal generation on pre-2022 processors can lead to thermal throttling after prolonged sessions.

Setup Guide

How to Install & Run Gemma 4 E2B

Step-by-step setup guide to initialize and chat with the model inside OfflineGPT.

Open OfflineGPT on an Android device with at least 8 GB of RAM.
Tap the model selector in the top bar and locate Gemma-4-E2B-it under the Multimodal section.
Tap Download (~2.58 GB). The download proceeds via Android's background download service.
Once completed, select Gemma 4 E2B. You can now type, attach photos via the paperclip icon, or record voice prompts via the microphone.
Optional: Open model settings to select OpenCL GPU acceleration or adjust context token limits.

FAQ

Frequently Asked Questions

Verified details on memory use, multi-modal features, and device support.

How does Gemma 4 E2B compare to Gemma 4 E4B?

Gemma 4 E2B (~2.58 GB) is Google's lightweight multimodal model optimized for low RAM consumption and fast generation on standard 8 GB Android devices. Gemma 4 E4B (~3.66 GB) has twice the parameters, providing deeper reasoning and higher visual fidelity for devices with 8 GB to 12 GB+ RAM. Both models run natively on Android via LiteRT-LM.

Does Gemma 4 E2B send images or audio to Google?

No. The LiteRT-LM engine runs locally on your phone's processor and GPU. Image pixels and audio waveforms are processed entirely in memory on the device. Data only leaves your phone if you explicitly toggle on the optional web search tool.

Can I use Gemma 4 E2B in airplane mode?

Yes. Once downloaded, all text, image, and audio understanding features work 100% offline without an active network connection.

Explore More

Related On-Device Models

Compare with other verified models supported in OfflineGPT.