No persistent endpoint
You find a GPU, quantize an 8B-parameter model and keep it running yourself.
NAIC 2026 · Developer Infrastructure
N-ATLaS, one import away.
OpenAtlas is a single TypeScript SDK for N-ATLaS. Call Nigeria's sovereign LLM and its four local-language speech models without a GPU, a quantization step or five separate interfaces.
$ npm install @openatlas/sdk
import { OpenAtlas } from "openatlas"; const client = new OpenAtlas({ apiKey: process.env.OPENATLAS_API_KEY }); const response = await client.chat({ messages: [{ role: "user", content: "Ṣe o le ṣàlàyé ìdí tí ọ̀run fi jẹ́ búlúù?" }], user: "your-end-user-id", // required: one ID per end user }); console.log(response.content); // → { content: string, model: "NCAIR1/N-ATLaS", attribution: "Powered by Awarri" }
One client. Four language codes.
ha Hausa
yo Yoruba
ig Igbo
en-ng Nigerian-accented English
The gap
A model is not an ecosystem.
N-ATLaS is open and downloadable. Between the Hugging Face page and a working app sits infrastructure work every developer ends up repeating.
01
02
No single client library
One LLM and four ASR checkpoints, each with its own interface, for you to stitch together.
03
No starter templates
Nothing to copy for the obvious first builds: a citizen chatbot, a classroom tool, a customer-service line.
What OpenAtlas is
The missing layer between the model and your app.
Not a new model, and not another app on top of N-ATLaS. Three parts that make building on it fast.
One client for five models.
A single npm package routes chat() to the LLM and transcribe() to the right ASR model by language code, with typed errors, retries and timeouts. It also repairs broken Nigerian-language text with normalizeText() and sends corrections back with reportIssue().
Method surface
- chat({ messages, user, language? })
- transcribe({ audio, language, user })
- normalizeText(text, { language? })
- reportIssue({ kind, output, correction })
- speak({ text, language }) — stretch
Why TypeScript first
It matches the stack of the starter kits and the gateway, and ships the inference settings already tested against the real N-ATLaS weights. A Python port is a roadmap item, not part of this submission.
Design principle
One consistent interface, whichever N-ATLaS model actually answers.
Three runnable reference apps.
Citizen services, education and customer service. Each is a minimal page or script that sends a real request through the SDK to the hosted endpoint and back.
What each proves
- Citizen Services — normalizeText() and chat() over a small demo civic dataset
- Education — chat() reframed by level and language, with reportIssue() for wrong answers
- Customer Service — transcribe() chained into chat(), with transcript corrections via reportIssue()
Run one
npm install && npm start, pointed at the live endpoint with a single environment variable.
Labeled honestly
Demo datasets and shortcuts are marked as demo inside the interface itself.
One gateway in front of N-ATLaS. No GPU on your side.
The OpenAtlas gateway issues your key, counts active users against the licence cap and forwards each call to a GPU host serving the N-ATLaS LLM and all four ASR models. You never provision hardware or manage weights just to try the model.
What's served
- N-ATLaS LLM — a Llama-3 8B fine-tune, quantized as needed for the GPU tier
- ASR for Hausa, Yoruba, Igbo and Nigerian-accented English
- One base URL, a distinct endpoint path for each model
Who it's for
Nigerian and West African developers, academic and research teams, and anyone building civic, education or support tools in local languages.
Honest scope
New infrastructure built for this submission. Not load-tested, with no uptime or SLA claim. The GPU host can change behind the gateway without any change to your code, and first-call latency while models load is documented up front.
The API
Five methods. That's the whole surface.
Request and response shapes echo familiar chat-completion conventions to lower the learning curve. Every model call except the optional speech one is served by an N-ATLaS model.
client.chat()
Text reasoning
Messages in, N-ATLaS LLM completion out. Optional language hint.
POST /v1/chat/completions
client.transcribe()
Speech to text
Audio plus a language code, routed to the matching N-ATLaS ASR model.
POST /v1/audio/transcriptions
normalizeText()
Text repair
Fixes Nigerian-language characters broken by typing or scraping, such as Æ™asa to ƙasa. Runs locally. It does not add tone marks that were never typed.
Local · no request
client.reportIssue()
Corrections
Flag a wrong N-ATLaS output with its correction. Every app becomes an opt-in source of corrected local-language data.
POST /v1/issues
client.speak()
Optional
Audio rendering
Final-stage rendering of text N-ATLaS already produced, by a separate speech model. Clear in English; Hausa, Yoruba and Igbo are experimental.
POST /v1/audio/speech
How it flows
Show the pipeline. Don't hide it.
Every starter kit displays each intermediate step, such as the transcript before the reply, so you can see N-ATLaS doing the work.
Customer Service
InputVoice note
transcribe()N-ATLaS ASR, by language
normalizeText()Repair broken characters
chat()N-ATLaS LLM drafts reply
speak() · stretchAudio reply
Citizen Services
InputQuestion in text
normalizeText()Repair broken characters
chat()N-ATLaS LLM + civic context
OutputResponse shown
Education
InputQuestion in a local language
chat()N-ATLaS LLM + instructional framing
OutputExplanation shown
Built on N-ATLaS
What's in the critical path
NAIC requires genuine N-ATLaS integration. OpenAtlas has no general-purpose model anywhere in the reasoning or transcription path.
In the path
- The N-ATLaS LLM handles every reasoning call
- N-ATLaS ASR handles every transcription
- Text-to-speech, if shipped, only renders audio from text N-ATLaS already wrote
- Demos run against the live endpoint, never mocks
Out of scope
- A general-purpose model standing in for N-ATLaS
- Fine-tuning or retraining any N-ATLaS model
- Commercial positioning, given the model's non-commercial licence
- Uptime, SLA or multi-tenant scale claims
Starter kits
Pick a niche. Swap in your content.
Reference implementations, not finished products. No persistence layer, no auth, no dataset beyond a labeled demo fixture.
01
Citizen Services
Local-language Q&A over a small civic-info dataset, such as how to renew an ID.
normalizeText() → chat()
02
Education
Tutoring-style explanations in Hausa, Yoruba or Igbo, reframed by level.
chat() · reportIssue()
03
Customer Service
A voice note is transcribed, then a reply is drafted for the support team.
transcribe() → normalizeText() → chat()
Known limitations
Read this before you build.
Stated plainly, not buried in a footnote.
First-call latency
The first call after the backend starts takes longer while the models load. Expect it, and retry if a request times out.
Non-commercial licence
N-ATLaS caps usage at 1,000 active users per 30 days and is non-commercial. The hosted endpoint is a developer and research resource, not a production SaaS.
No SLA
This is new infrastructure and it has not been load-tested. No uptime or reliability claim is made.
Speech output is a stretch
speak() depends on NAIC confirming the pairing is allowed. If not, it is dropped without weakening the core SDK.
No dashboard yet
API keys are requested through a short form and issued by hand. There is no usage analytics or billing UI.
Text repair, not restoration
normalizeText() fixes characters that were corrupted. It does not add tone marks that were never typed; N-ATLaS-based tone restoration is a roadmap item.
A real N-ATLaS response in five minutes.
That is the bar the quickstart is held to: copy, paste, run.