Source code of HAYAT v1.0.0, mirrored from GitHub. Models: HAYAT-35B-A3B-Uncensored-GGUF and HAYAT-Open-4B. Talk to the creator in GitHub Discussions.
HAYAT
A private, local-first AI assistant for Windows — your own AI that runs on your own computer.
Talk to it, let it research the web, see through your camera and screen, use your apps, write code, make images and video, track markets and jobs, and remember what matters — with a local language model and no cloud account.
Quick start · What it can do · HAYAT explained · How to use it · Roadmap · Contributing · Discussions · Models
Why HAYAT
- Private by default. The language model, memory, files and camera stay on your machine. Nothing leaves it unless you ask. Web reading can go through Tor.
- Zero-cost. No paid APIs, no subscriptions. Paid fallbacks are blocked by policy, not by promise.
- It acts, and proves it. HAYAT runs real tools — files, browser, Windows apps, research, media — and only says something was done when the tool result shows it. Anything that touches the outside world waits for your approval.
- It tells the truth about itself. It says when it does not know, when a step failed and why, and cites the pages it read. Answers are plain language, never raw data dumps.
- One assistant, many surfaces. Desktop app, voice, terminal and an authenticated web client share one runtime.
What it can do
| Area | Highlights |
|---|---|
| Talk and think | Natural chat with a local 35B mixture-of-experts model; deep reasoning mode; corrections remembered |
| Voice | Push-to-talk or hands-free, local speech recognition and speech output |
| Research | Background web research with cited sources, freshness and conflict checks; optional Tor and onion research |
| See | Webcam and screen understanding, text reading (OCR), person detection, consent-only face enrollment |
| Use your computer | Open and drive Windows apps by control name, browse websites, fill forms with your OK |
| Create | Images (SDXL) and short local video clips, documents, PDFs and Word files, code |
| Remember | Long-term memory with recall, a running picture of what you are working on, and lessons from past mistakes |
| Work and money | Job search with CV matching, business planning, invoices, a read-only Trade Center for crypto and stocks |
| World awareness | Global intelligence: public cameras, hazards, weather, satellites and news on a live map |
| Privacy tools | Tor masking, location-exposure checks, network guard, secret redaction, Windows Hello owner checks |
| Build | Reverse-engineering workbench, code tools, automations and reminders, worker teams for bigger jobs |
Deep dive: HAYAT explained covers how every part works inside — the 21-part brain, memory and its "DNA" association graph, 24/7 learning and self-upgrade, predictions and their measured track records, people and client search, jobs, markets, global intelligence, vision, media, voice, privacy and limits — with today's test results. The full list of 40 areas, each with words to say and the window to open, is in How to use HAYAT.
Feature status
| Capability | Status | Notes |
|---|---|---|
| Conversation and agent loop | Working | One shared core for text and voice |
| Public-web research | Working | Free provider chain; web content is treated as untrusted data |
| Windows and browser control | Working | Permissioned, observed and verified |
| Voice | Working | Local speech in and out |
| Camera OCR | Working | Windows OCR first, bundled Tesseract as fallback |
| Image and video generation | Working | Local stable-diffusion.cpp engine on your GPU |
| Tor web reading | Working | Remembered on/off; starts on demand; never silently falls back to direct |
| Fine-tuning | Gated | Local QLoRA path exists; promotion requires passing held-out evaluation |
Quick start
You need: Windows 10/11, Python 3.11 or newer (developed and tested on 3.14), Ollama, about 30 GB free disk for the default model, and ideally an NVIDIA GPU with 8 GB VRAM (CPU works, slower; RAM 32 GB+ recommended for the 35B model).
# 1. Get the code
git clone https://github.com/mohsinkhanshaik/HAYAT.git
cd HAYAT
# 2. Install Ollama and HAYAT's brain from Hugging Face (22 GB, includes vision)
winget install Ollama.Ollama
ollama pull hf.co/Mohsinkhanshaik/HAYAT-35B-A3B-Uncensored-GGUF:Q4_K_M
ollama cp hf.co/Mohsinkhanshaik/HAYAT-35B-A3B-Uncensored-GGUF:Q4_K_M qwen3.6-35b-a3b-uncensored:q4_k_m
# 3. Set up and start
.\setup.ps1 # virtual environment, dependencies, embeddings, and the hayat model
python run.py # opens the desktop app
On a smaller machine, use the 4B model instead:
ollama pull hf.co/Mohsinkhanshaik/HAYAT-Open-4B:Q4_K_M and select it in Models. Run
python scripts/doctor.py any time to see what is installed and what is missing. More detail:
getting started and troubleshooting.
Models
| Model | Size | Use |
|---|---|---|
| HAYAT-35B-A3B-Uncensored-GGUF | 22 GB | The default brain (Qwen3.6-35B-A3B, about 3B active per token), with vision |
| HAYAT-Open-4B | 2.5–4.3 GB | Small experimental model for lighter machines |
HAYAT works with any Ollama model; see models.
How it is built
- Python desktop app (Tkinter), local Ollama for language and vision, stable-diffusion.cpp for media, Playwright for browsing and Windows UI Automation for apps.
- 597 modules, about 166,000 lines, 7,865 automated tests (7,738 run offline in about 20 minutes; 127 opt-in live checks against real models and networks).
- Every tool is registered with a permission class; every action is checked against its result before HAYAT claims it.
- Start reading at architecture and the technical tour.
Help make it more powerful
HAYAT is built in the open, and the next steps are clear. The roadmap lists them, each with an open issue:
- More honest, more uncensored answers — measure false refusals and unrequested disclaimers, and evaluate newer uncensored local models (#1, #2).
- Deeper reasoning — verify multi-step answers before replying (#3).
- Faster and broader — quicker start, more free research sources, Linux support, a good default for small machines (#5–#7, #10).
- First contribution? Start with a good first issue.
Read CONTRIBUTING.md for the layout, how to run the tests, and the few rules every change follows (code + tests + docs, honest results, no paid services). Security problems: see SECURITY.md.
Talk to the creator
Questions, ideas, show-and-tell, or want to work on HAYAT together? Start a thread in Discussions — that is the best way to reach me. Bugs go in Issues.
Free, forever, and yours to run
HAYAT runs entirely on your computer. There is no server to depend on and nobody else's machine needs to be on: once installed, your copy works on its own. It costs nothing — no account, no subscription, no paid API; paid services are blocked by policy. GitHub and Hugging Face downloads are free.
Licence
HAYAT is free software under the GNU AGPL-3.0: use it, study it, change it and share it. If you distribute a changed version, or offer it to others over a network, share your source under the same licence. Two additional terms in NOTICE keep its origin visible: every copy keeps the credit "HAYAT, created by Mohsin Khan Shaik", and changed versions say they are based on HAYAT rather than presenting themselves as the original. Third-party components keep their own licences — see NOTICE.
HAYAT was created by Mohsin Khan Shaik.