N-ATLAS Kit

Developer toolkit for N-ATLaS, Nigeria's sovereign multilingual model, and the four NCAIR1 speech models that sit beside it.

N-ATLaS speaks Hausa, Igbo, Yorùbá, and Nigerian English. The playground also offers Nigerian Pidgin (beta). The speech models are separate Whisper Small checkpoints, one per language, trained on voice recorded across Nigeria's six geopolitical zones. This repository is the client and the self-hosting kit. It does not ship weights.

A Hausa draft of this page is at Bayani. It has not been reviewed by a native speaker.

See it in action

Two recordings from the playground.

Hausa chat in the playground, then Get code copies a working SDK call.

Yorùbá chat and the Tune panel, then Speech transcribes audio and translates it.

Why this exists

The weights are on Hugging Face. Using them still means accepting gated terms, downloading about 15 GB, learning that the chat template needs a date_string, finding four separate ASR repositories, and discovering that each one only accepts 30 seconds of 16 kHz mono audio.

N-ATLAS Kit is the layer in front of that:

Piece	What it is
serve/	One OpenAI-compatible gateway for chat and all four speech models
n-atlas	JavaScript / TypeScript SDK. npm install n-atlas or pnpm add n-atlas
natlas	Python SDK, sync and async. pip install natlas
Examples	Small apps you can run against a gateway
Playground	Live browser app: chat, speech, and get-code

There is no public NCAIR inference API in this release. Every client takes a baseURL. The hosted gateway used for the National AI Innovation Challenge demo is:

https://kambah123--natlas-serve-natlasservice-serve.modal.run

GET /health on that host does not need an API key. It scales to zero, so the first request after an idle period is a cold start (the container reloads the weights). Chat and transcription need the Bearer key stored in the Modal secret natlas-api. That key is not in this repository.

Ten minutes, if a gateway is already up

1. Install Node.js 20 or newer, or Python 3.10 or newer.
2. Set NATLAS_BASE_URL to the gateway origin (with or without /v1) and NATLAS_API_KEY to a key that gateway accepts.
3. Follow the quickstart for one chat call and one transcription, or run the voice-note translator.

If those variables are missing, the SDKs and the example apps stop and say so. They do not print a made-up transcript or reply.

What the kit will not do

• It will not call GPT, Claude, Gemini, or any other vendor's model. A model id that does not start with NCAIR1/ is rejected in the client.
• It will not download weights for you unless you self-host. The HTTP SDKs only speak to a gateway.
• It does not include a reviewed Hausa translation of the whole site. Overview and quickstart exist in Hausa and are marked as needing native-speaker review.
• Benchmark numbers are not invented here. GPU sizing in the self-hosting guide is the planning arithmetic already written in serve/README.md, from the model card and config.json.

Where to go next

• Quickstart — install, environment, first chat, first transcription
• Gateway API — the HTTP surface serve/ actually implements
• JavaScript SDK and Python SDK
• Self-hosting — Modal, Docker Compose, Hugging Face terms, GPUs
• Limits and troubleshooting
• Licensing and attribution
• How the kit uses N-ATLaS