--- license: apache-2.0 tags: - pocketllm - on-device - edge-ai - mobile - android - offline - mediapipe - gguf - llama - gemma - qwen - phi language: - en - hi - multilingual pipeline_tag: text-generation --- # PocketLLM Model Collection
**The official model collection for [PocketLLM](https://github.com/amareshhebbar/PocketLLM)** *Private • Offline • On-Device AI for Android* [![GitHub](https://img.shields.io/badge/GitHub-PocketLLM-black?logo=github)](https://github.com/amareshhebbar/PocketLLM) [![Platform](https://img.shields.io/badge/Platform-Android-green?logo=android)](https://github.com/amareshhebbar/PocketLLM) [![License](https://img.shields.io/badge/License-Apache%202.0-blue)](LICENSE)
--- ## What is PocketLLM? PocketLLM is an Android app that runs large language models **completely on your device** — no internet required during inference, no data sent to the cloud, no subscriptions. This repository hosts the curated model collection optimized for mobile edge inference using: - **MediaPipe LLM Inference** (for `.bin` and `.task` files — Gemma family) - **llama.cpp via llama.rn** (for `.gguf` files — Llama, Phi, Qwen, SmolLM, Gemma 3) --- ## Model Catalog ### ⚡ Ultra Fast — Under 1 GB, fits any Android phone | Model | File | Size | Format | RAM | |-------|------|------|--------|-----| | Qwen 2.5 0.5B | `qwen2.5-0.5b-instruct-q4_k_m.gguf` | 0.4 GB | GGUF | 2 GB | | Gemma 3 1B | `gemma-3-1b-it-q4_k_m.gguf` | 0.7 GB | GGUF | 2 GB | | Llama 3.2 1B | `llama-3.2-1b-instruct-q4_k_m.gguf` | 0.8 GB | GGUF | 3 GB | | SmolLM2 1.7B | `smollm2-1.7b-instruct-q4_k_m.gguf` | 1.0 GB | GGUF | 3 GB | ### ⚖️ Balanced — 1–2 GB, great quality on mid-range phones | Model | File | Size | Format | RAM | |-------|------|------|--------|-----| | Gemma 1.1 2B (CPU) | `gemma-1.1-2b-it-cpu-int4.bin` | 1.35 GB | MediaPipe | 4 GB | | Gemma 1.1 2B (GPU) | `gemma-1.1-2b-it-gpu-int4.bin` | 1.35 GB | MediaPipe | 4 GB | | Llama 3.2 3B | `llama-3.2-3b-instruct-q4_k_m.gguf` | 2.0 GB | GGUF | 4 GB | ### 🚀 Powerful — Best quality, needs 5GB+ RAM | Model | File | Size | Format | RAM | |-------|------|------|--------|-----| | Phi-3.5 Mini | `phi-3.5-mini-instruct-q4_k_m.gguf` | 2.4 GB | GGUF | 5 GB | | Gemma 3 4B | `gemma-3-4b-it-q4_k_m.gguf` | 2.8 GB | GGUF | 6 GB | --- ## How to Use in PocketLLM Models are downloaded directly inside the PocketLLM app. Open the **Model Store** tab, select a model, and tap Download. The app handles everything automatically. **Direct download URLs:** ``` https://huggingface.co/AmareshHebbar/pocketllm-models/resolve/main/ ``` --- ## Fine-Tuning Roadmap > *Coming soon — within the next 2 weeks* We are fine-tuning these base models specifically for **on-device conversational AI** on mobile: ### Goals - Better persona adherence (the model stays in character consistently) - Shorter, more natural responses (base models tend to be verbose on mobile) - Improved memory utilization (uses injected memories naturally) - Better instruction following for small context windows (2048 tokens) - Hindi + English code-switching support (for India market) ### Training approach - **Base:** Llama 3.2 1B, Gemma 3 1B (smallest models first — fastest iteration) - **Method:** QLoRA fine-tuning (4-bit, similar to the AxioMapper training setup) - **Dataset:** Curated conversational dataset optimized for mobile edge constraints - **Hardware:** Single A100 via RunPod - **Framework:** Unsloth (2x faster fine-tuning, same as AxioMapper) ### Fine-tuned models (coming soon) - `pocketllm-llama-1b-v1.gguf` — Llama 3.2 1B fine-tuned for mobile chat - `pocketllm-gemma-1b-v1.gguf` — Gemma 3 1B fine-tuned for persona consistency --- ## Model Selection Guide ``` Your phone has... Best model to start with ───────────────────────────────────────────────────── 2 GB RAM (budget) → Qwen 2.5 0.5B (fastest) 3 GB RAM → Llama 3.2 1B (balanced speed) 4 GB RAM (mid-range) → Gemma 1.1 2B (best all-rounder) 6 GB RAM → Llama 3.2 3B (better quality) 8 GB RAM (flagship) → Phi-3.5 Mini (best for coding) ``` --- ## Technical Details ### MediaPipe format (`.bin`, `.task`) - Used for Gemma family - Runs via Google's MediaPipe LLM Inference SDK - Supports CPU and GPU acceleration (Vulkan/OpenCL) - Integrated via `react-native-llm-mediapipe` ### GGUF format (`.gguf`) - Used for Llama, Qwen, Phi, SmolLM, Gemma 3 - Runs via llama.cpp (the gold standard for mobile inference) - Q4_K_M quantization — best balance of size and quality - Integrated via `llama.rn` --- ## Repository Structure ``` pocketllm-models/ ├── gemma-1.1-2b-it-cpu-int4.bin ← Gemma 2B CPU (MediaPipe) ├── gemma-1.1-2b-it-gpu-int4.bin ← Gemma 2B GPU (MediaPipe) ├── gemma-3-1b-it-q4_k_m.gguf ← Gemma 3 1B (GGUF) ├── gemma-3-4b-it-q4_k_m.gguf ← Gemma 3 4B (GGUF) ├── llama-3.2-1b-instruct-q4_k_m.gguf ← Llama 3.2 1B (GGUF) ├── llama-3.2-3b-instruct-q4_k_m.gguf ← Llama 3.2 3B (GGUF) ├── phi-3.5-mini-instruct-q4_k_m.gguf ← Phi-3.5 Mini (GGUF) ├── qwen2.5-0.5b-instruct-q4_k_m.gguf ← Qwen 2.5 0.5B (GGUF) └── smollm2-1.7b-instruct-q4_k_m.gguf ← SmolLM2 1.7B (GGUF) ``` --- ## License - **Gemma models:** [Gemma Terms of Use](https://ai.google.dev/gemma/terms) - **Llama models:** [Meta Llama 3.2 Community License](https://llama.meta.com/llama3_2/license/) - **Phi models:** [MIT License](https://huggingface.co/microsoft/Phi-3.5-mini-instruct/blob/main/LICENSE) - **Qwen models:** [Apache 2.0](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct/blob/main/LICENSE) - **SmolLM2:** [Apache 2.0](https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct/blob/main/LICENSE) - **Fine-tuned models by PocketLLM:** Apache 2.0 --- ## Built By **Amaresh Hebbar** Building PocketLLM: the only mobile app that runs a full AI agent stack — smart routing, persona memory, MCP tools — completely offline on Android. *"Your AI. Your phone. Nobody else's business."* [![GitHub](https://img.shields.io/badge/GitHub-amareshhebbar-black?logo=github)](https://github.com/amareshhebbar)