Welcome to Haven AI Companion 🌸

Haven is a fully self-hosted, offline-first, private AI companion and smart assistant ecosystem. We focus on optimizing state-of-the-art Large Language Models (LLMs) and Speech models to run locally on commodity hardware, giving users absolute ownership over their data, companion personalities, and voice pipelines.

This Hugging Face organization hosts our custom quantizations, character card datasets, and local AI model packages designed for the Haven ecosystem.

📦 What We Host

1. Custom Quantized Models (GGUFs)

We quantize state-of-the-art open models (such as the Google Gemma 4 and Qwen families) utilizing advanced Importance Matrix (imatrix) calibration:

• Optimal Compression: We target quantizations like IQ4_XS and Q3_K_S to shrink model sizes by up to 50% while preserving maximum cognitive coherence.
• CPU Optimized: Models are structured to execute with low-latency on standard consumer laptops, home servers, and mobile devices without requiring expensive GPUs.

2. Multi-Character Configurations (Companion Cards)

We compile and distribute companion profile assets (standard Tavern V1/V2 PNG character cards with embedded system prompts, personalities, first messages, and dialogue examples) that can be imported directly into the Haven Mobile client.

🏗️ Ecosystem Architecture

• Haven Server (GitHub): Our low-latency C# orchestrator that manages companion profiles, handles text/voice WebSockets, split-prompt parsing, and local TTS pipelines.
• Haven Mobile (GitHub): Our native Android application written in Kotlin and Jetpack Compose, featuring local Room database caching and ONNX-driven Voice Activity Detection (VAD) for hands-free calling.

🛠️ Quick Setup with Ollama

To run our custom quantized models within your local Ollama setup:

1. Download the GGUF file from our model repository (e.g., ssfdre38/gemma4-e4b-merged-iq4xs-turbo).
2. Create a Modelfile in the same directory:

   FROM ./gemma4-e4b-merged-iq4xs-turbo.gguf
   PARAMETER num_ctx 16384
   TEMPLATE """<start_of_turn>user
   {{ .Prompt }}<end_of_turn>
   <start_of_turn>model
   {{ .Response }}<end_of_turn>"""
   

3. Build and launch your model:

   ollama create gemma4-turbo -f Modelfile
   ollama run gemma4-turbo
   

Keep your AI local, secure, and private.