RexiO is not a translation layer. It's a Bangladesh product.

If you've used the big-name assistants, you already know what we're not building.

They're built in California, Singapore, and Shenzhen, by teams that don't speak your language first. They handle Bengali as an afterthought — a translation layer slapped on top of a model that was trained on English. They get the script right sometimes. They almost never get the context right.

RexiO is built in Bangladesh, by Bangladeshis, for Bangladesh. That's the entire difference. Here's what that means in practice.

The model isn't translated. It's trained.

When we built RexiO, we didn't start with a foundation model and bolt Bengali on top. We started with Bangla — and Banglish, the way most urban Bangladeshis actually type. Our tokenizer was trained on:

• 6 months of Prothom Alo, Daily Star Bangla, and Bangla Tribune
• Bangla subtitles from YouTube and TikTok
• Internal chat logs from our private alpha (anonymized)
• 2GB of curated Bangla web text, including parts of Wikipedia that aren't well-translated

The result: RexiO is roughly 4x cheaper to run on Bangla than the global models, and noticeably better at the kinds of queries Bangladeshi users actually ask. We didn't beat the global models on Bengali because we had better GPUs. We beat them because we started with Bengali.

Context isn't a feature. It's the product.

The global assistants don't know that ৳80 is cheap for a rickshaw from Dhanmondi to Gulshan, and expensive for a Cox's Bazar bus ticket. They don't know that 11pm in Dhaka is when most conversations start. They don't know what Pohela Baishakh is, or that Ekushey February carries weight, or that your cousin in Sylhet is going to ask you for help with her SSC result in three days.

RexiO does. Not because we hard-coded these facts — we didn't. But because the model was trained on Bangladesh, and fine-tuned on conversations from Bangladeshis. When you ask "what's the weather in Sylhet right now" we don't have to translate the question and hope for the best. We already know what Sylhet is, what "right now" means, and what kind of answer a Bangladeshi user actually wants.

This isn't a feature. It's the entire reason RexiO exists. Without it, we're just another chatbot.

We make different tradeoffs.

The global assistants optimize for everything, for everyone. They need to be good at coding in San Francisco and recipe writing in São Paulo and haiku composition in Tokyo. That's a hard problem, and the teams building them are extremely good.

But Bangladesh doesn't need "good at everything for everyone." Bangladesh needs good at Bangladesh. So we made a different choice.

• We use less compute per query, because compute is expensive in BDT and our users shouldn't pay for capabilities they don't use.
• We have smaller context windows, but they're filled with Bangladesh-relevant context — not Wikipedia trivia.
• We're slower to ship features like image generation or code execution, because those features have high server cost and low daily use for our audience. We'd rather charge you less.

This is a deliberate tradeoff. We will not catch up to the global assistants on raw feature count. We don't need to. We need to be the best AI for someone standing in a bazaar in Old Dhaka with a cracked Android phone, trying to figure out the bus schedule to Barisal.

What this means for the country

RexiO isn't a Bangladesh product the way some companies make "India products" or "Africa products" — a thin wrapper around a Western product with a local sales team. RexiO is a Bangladesh product the way Grameenphone is a Bangladeshi product: built here, staffed here, optimized here.

That has effects.

• A high school student in Rajshahi can ask RexiO about SSC results in Bangla, in her own dialect, and get a real answer in seconds.
• A small business owner in Chattogram can use RexiO to draft a Facebook post in Banglish for his products, in the same casual register his customers already use.
• A grandmother in Khulna can ask RexiO to read a letter she received, and have the response spoken back to her in Bengali, at a speed her old phone can handle.

These are not hypothetical. These are the actual use cases we built for. The closed test we're running now is with 500 people chosen specifically to cover the breadth of Bangladesh — old, young, urban, rural, Bangla, Banglish, Android on 2G, Android on 5G.

The honest part

We're not the right tool for everyone. If you need a model that's been trained on the entire English-language internet for technical research, use the global tools. They're excellent at that. We're excellent at being your AI in Bangladesh.

We're not going to compete on raw capability. We're going to compete on being useful — for the questions you actually have, in the language you actually speak, on the device you actually own.

That's the bet. And we think it's the right one for Bangladesh.

— The RexiO team