HUMAIN Chat
Built in Arabic. Designed for the world.
If you speak Arabic, you've probably noticed most AI tools treat it as a translation layer bolted onto an English-first model. They stumble on dialects, miss cultural context, and default to formal Modern Standard Arabic that nobody actually speaks day to day. For 400M+ Arabic speakers, that's a pretty broken experience.
So we built HUMAIN Chat — an assistant that's Arabic-first by design but fully bilingual in English. It runs on ALLaM 34B, a model we trained in Saudi Arabia on one of the largest Arabic datasets assembled to date, then refined with 600+ domain experts and 250 evaluators to get the dialectal and cultural nuance right.
A few things we're proud of:
Dialects: handling everything from MSA to regional spoken Arabic in one model, including voice input across dialects.
Bilingual switching: you can flip between Arabic and English mid-conversation and it keeps up.
Sovereign hosting: the whole thing runs on our own infrastructure inside the Kingdom, fully PDPL-compliant. That mattered a lot for trust here.
We launched first in Saudi Arabia (web, iOS, Android) and are rolling out across the region next, then wider.
The most interesting part has been building for a market the big labs deprioritize — lots of demand, very little tailored supply.
if you've built for a non-English or underserved-language market, what has caught you off guard most; the data, the evaluation, or the go-to-market?
1 Comment
I think one of the biggest mistakes in multilingual AI is treating language as vocabulary instead of context. Dialects carry different assumptions, cultural references, and ways of expressing intent, so simply translating an English-first model rarely feels native. Building around those nuances instead of adding Arabic as another supported language feels like a much stronger strategic position.