~2.58 GB
Gemma 4 E2B
By Google · LiteRT-LM (.litertlm)
Google's multimodal edge model in official .litertlm format. Understands text, image attachments, and spoken audio natively in a single prompt.
TextVision (Images)Audio Input
On-Device AI Model Catalog.
Verified open-weight AI models optimized for offline Android inference in OfflineGPT. Check verified RAM requirements, inference engines, and download sizes before installing.
llama.cpp Engine
GGUF Quantization CPU multi-threading and GPU offload for Llama 3.2, Qwen3 Thinking, and fast text instruct models.
LiteRT-LM Engine
Google AI Edge Native multimodal understanding of text, images, and audio attachments in official .litertlm format.
MNN & QNN Engine
Hexagon NPU Acceleration On-device vision understanding and Stable Diffusion image generation accelerated on Snapdragon chips.
Deep-Dive Guides
Verified On-Device Foundation Models
In-depth technical specifications, hardware requirements, and setup guides for the core on-device models.
~3.66 GB
Gemma 4 E4B
By Google · LiteRT-LM (.litertlm)
Google's 4-billion parameter multimodal model in .litertlm format. Delivers enhanced cognitive reasoning and high-fidelity visual understanding for devices with 8 GB to 12 GB RAM.
TextVision (Images)Audio Input
~2.60 GB
Qwen3 4B Thinking
By Alibaba Cloud / Qwen · GGUF
Dual-mode reasoning model with visible thinking stream. Generates step-by-step reasoning traces in real time before delivering the final answer.
Text
~1.42 GB
Llama 3.2 1B
By Meta · GGUF
Meta's lightweight mobile model in high-fidelity 8-bit quantization. Fast generation, low RAM usage, and instant response times on almost any Android device.
Text
Multimodal & Vision
Text, Image, and Audio Understanding
Models that inspect camera captures, photo attachments, or audio recordings directly on device.
~2.58 GB
Gemma 4 E2B
By Google · LiteRT-LM (.litertlm)
Google's multimodal edge model in official .litertlm format. Understands text, image attachments, and spoken audio natively in a single prompt.
TextVision (Images)Audio Input
~3.66 GB
Gemma 4 E4B
By Google · LiteRT-LM (.litertlm)
Google's 4-billion parameter multimodal model in .litertlm format. Delivers enhanced cognitive reasoning and high-fidelity visual understanding for devices with 8 GB to 12 GB RAM.
TextVision (Images)Audio Input
~888 MB
InternVL2.5 1B Vision
By OpenGVLab · MNN
Lightweight on-device vision model running via Alibaba's MNN framework. Capable of optical character recognition and image descriptions on 4 GB RAM.
TextVision (Images)
~2.48 GB
SmolVLM2 2.2B Vision
By Hugging Face · MNN
Detailed visual question answering and image understanding for Android devices with 6 GB or more of RAM.
TextVision (Images)
Reasoning
Visible Thinking Models
Models that stream their intermediate reasoning steps before answering, allowing verification of logical deductions.
~2.60 GB
Qwen3 4B Thinking
By Alibaba Cloud / Qwen · GGUF
Dual-mode reasoning model with visible thinking stream. Generates step-by-step reasoning traces in real time before delivering the final answer.
Text
Text & Writing
Fast On-Device Instruct Models
Compact open-weight models optimized for fast generation, low battery consumption, and phones with 4 GB to 6 GB RAM.
~1.42 GB
Llama 3.2 1B
By Meta · GGUF
Meta's lightweight mobile model in high-fidelity 8-bit quantization. Fast generation, low RAM usage, and instant response times on almost any Android device.
Text
~731 MB
LFM2.5 1.2B
By LiquidAI · GGUF
Liquid Neural Network architecture optimized for on-device inference. Smallest text model in the catalog with rapid generation and minimal battery drain.
Text
~1.85 GB
Gemma 2 2B
By Google · GGUF
Google's 2B parameter instruct model in Q5_K_M precision. High factual knowledge and clear technical explanations for mid-range phones.
Text
~2.10 GB
Qwen 2.5 3B
By Alibaba · GGUF
Versatile general-purpose model with strong multilingual fluency, coding assistance, and structured markdown outputs.
Text
~3.15 GB
Rocket 3B
By Community · GGUF
Community fine-tuned conversational model in 8-bit precision. Offers rich conversational responses for devices with 8 GB RAM.
Text
Diffusion & Speech
Local Image Generation & Speech Recognition
Qualcomm Snapdragon NPU-accelerated diffusion and offline Whisper speech-to-text running entirely on device.
~1.0 – 1.25 GB per style
Stable Diffusion (MNN + Hexagon NPU)
By Stability AI / Community · MNN
On-device image generation accelerated by Qualcomm Snapdragon NPU. Generates 512×512 artwork locally in anime or photorealistic styles without cloud tokens.
Text to Image
~78 MB (Base) / ~142 MB (Enhanced)
Whisper Voice Input
By OpenAI / whisper.cpp · GGUF
High-accuracy offline speech recognition. Transcribes voice input directly on device with zero audio uploaded to any server.
Audio to Text
RAM & Device Compatibility Guide
Running AI models locally requires dedicated memory. A phone with 4 GB RAM comfortably runs 1B models (Llama 3.2, LFM 2.5), 6 GB RAM enables 3B–4B thinking models, and 8 GB RAM is required for Gemma 4 multimodal and Stable Diffusion. Check our hardware tier breakdown before downloading large model files.
Get OfflineGPT
OfflineGPT on Google Play.
OfflineGPT · offlinegpt.tech · Developed by CodeAlp · com.offlinegpt.app