SABI is a private, offline AI coworker that turns ideas into working software on the low-cost laptops Africa already has — no cloud required.
# SABI — The Offline AI Coworker
**On-device AI for the hardware Africa actually has.**
SABI is a private, fully offline AI coworker. It runs a single quantized model
on a standard 8 GB laptop — no cloud, no API fees, no internet — staying under a
strict 7 GB memory budget. It chats, plans, writes code, and can **take real
action** (create files, run commands) through a Claude-Code-style agent, with
your approval. Use it from a polished terminal interface, a full-screen TUI, or
a ChatGPT-style web app.
---
## Live demo
**Try it right now, no install, no clone:** **8554ef3b3ec8bc6c1f.gradio.live**
A hosted instance of SABI's Coding Assistant, running the same `sabi-v1`
model and `sabi-yoruba-llm` translation layer as the rest of this repo,
behind a public link — click it and start chatting immediately. It mirrors
`sabi serve`'s web UI: multi-conversation history, document upload (PDF,
Word, Excel, PowerPoint, CSV, HTML, images) with summarization, and
Auto/Think/Code modes, plus the **Reply in Yoruba** toggle (or just type in
Yoruba — it auto-detects) for the African-language bonus. It's deliberately
**text-only** — no file access, no shell execution — since this link is
public on the open internet, unlike `sabi serve` itself.
> **Note for reviewers.** This is a temporary tunnel (Gradio's free share
> link, best-effort for up to ~1 week) proxying to the author's own machine —
> it can go offline if that machine is off or the link expires. If it's down
> when you check, every command in Quickstart below reproduces
> the identical experience locally in a couple of minutes
> (`pip install "sabi-llm[full]" && sabi serve`), or contact the author to
> restart it.
---
## Table of contents
- Live demo
- What is SABI?
- Quickstart
- Installation, in detail
- Optional extras explained
- Getting the models
- The three interfaces
- Command reference
- Configuration
- How SABI works
- Benchmarking & the 7 GB budget
- Project structure
- Development
- Hardware t …