Language Brain · For AI companies
Make your model understand how people really talk.
Big models learn from the written web, so they misread dialects, slang and code-switching. We build what's missing, one language at a time, starting with Tamil, and you plug it in.
Input · Jaffna Tamil
நாளைக்கு கடைக்கு வாறியோ?
A generic model
Unclear verb, read as a statement about the shop.
With Language Brain
“Are you coming to the shop tomorrow?” · dialect: jaffna · register: casual
What you get
Everything below the top layer.
- Early access
Language Brain API
One call to understand dialect, register, script and code-switching before your agent answers.
How it works → - Live
Models trained from scratch
Nila is trained from scratch on Tamil, English and Tanglish and answers in spoken Tamil. Talk to it now.
Talk to Nila → - Live
Native-speaker evals
Uyir Bench: a private test set where every question is approved by two native speakers, so it can't leak into training data.
See the report card → - Early access
Consented speech and text
Real conversations and voice recordings from native speakers, given with explicit consent and reviewed by people.
How we collect →
Built for the language
Tamil costs almost half as many tokens.
General tokenizers chop Tamil into tiny pieces, so every call costs more and less of the conversation fits. Nila's tokenizer was built for Tamil and Tanglish from day one.
Tamil
45% fewer tokens
- GPT-4o tokenizer
- 3.24
- Nila
- 1.79
Tanglish
36% fewer tokens
- GPT-4o tokenizer
- 3.33
- Nila
- 2.13
Tokens per word on held-out text, measured by Uyir Lab with GPT-4o's o200k tokenizer.
Language Brain
A layer any AI agent plugs into.
Language Brain sits between your agent and the people it serves. It understands dialects, slang and cultural context, so the agent you already have becomes accurate in languages it used to get wrong.
- 1
01
Listen
Native speakers share real conversations, stories and voice recordings, with explicit consent and human review.
- 2
02
Learn
We build language models and evaluation sets from that data, starting with Nila, our Tamil model trained from scratch.
- 3
03
Plug in
Your agent calls Language Brain to understand people and reply the way they actually talk.
Input
Jaffna Tamil, casual. A generic model often misreads the dialect verb.
From the lab
Nila is training right now.
Nila 1B, our own model with a billion parameters, trained from scratch on 9.6 billion tokens of Tamil, English and Tanglish (25 billion tokens of training in all). These numbers come straight from the training run.
nila · live
Connecting to the training run…
Training loss
Languages
The languages big models miss.
We start where the gap is widest: languages with tens of millions of speakers and very little of the data big models learn from.
- தமிழ்Live
Tamil · 86M+ speakers
Spoken Tamil differs sharply from written Tamil, and dialects from Jaffna to Chennai barely appear online.
Visit Tamil → - සිංහලNext
Sinhala · 20M+ speakers
Written and spoken Sinhala are almost two different registers, and models trained on formal text sound stiff in conversation.
Join the waitlist → - پنجابی·ਪੰਜਾਬੀPlanned
Punjabi · 156M+ speakers
About 89M Punjabi speakers in Pakistan (2023 census) write in Shahmukhi, a Perso-Arabic script, but much of the language lives in speech, not text. Many speakers write online in Urdu or English instead.
Join the waitlist → - KiswahiliPlanned
Swahili · 97M+ speakers
Most of its speakers use it as a second language, and everyday speech mixes in English (Sheng in Nairobi), which formal web text misses.
Join the waitlist → - Harshen HausaPlanned
Hausa · 94M+ speakers
One of Africa's most spoken languages, yet with little web text, written in both Latin and Arabic (Ajami) script.
Join the waitlist → - አማርኛPlanned
Amharic · 60M+ speakers
Its own script (Ge'ez) and rich word forms get split into many pieces by tokenizers built for English.
Join the waitlist → - Basa JawaPlanned
Javanese · 68M+ speakers
Speech levels such as ngoko and krama change the words depending on who you are talking to, and many speakers write online in Indonesian instead.
Join the waitlist → - FulfuldePlanned
Fula · 39M+ speakers
Spoken across West and Central Africa and written in Latin, Arabic (Ajami) and its own Adlam script. Common Crawl's language detector does not even recognize it.
Join the waitlist → - Èdè YorùbáPlanned
Yoruba · 50M+ speakers
A tonal language where meaning rides on diacritics that most web text leaves out.
Join the waitlist → - भोजपुरीPlanned
Bhojpuri · 52M+ speakers
Mostly spoken, often written in Devanagari mixed with Hindi, and not recognized as a separate language by Common Crawl's language detector.
Join the waitlist → - Afaan OromooPlanned
Oromo · 45M+ speakers
45M+ native speakers, yet fewer than 1 in every 100,000 web pages is in Oromo.
Join the waitlist → - پښتوPlanned
Pashto · 55M+ speakers
Strong regional dialects and very little high-quality digital text.
Join the waitlist → - Asụsụ IgboPlanned
Igbo · 34M+ speakers
Tonal, with many dialects, and the little text online is often missing its tone marks.
Join the waitlist → - ភាសាខ្មែរPlanned
Khmer · 21M+ speakers
Written without spaces between words, so standard tokenizers and search struggle from the first step.
Join the waitlist → - Kreyòl ayisyenPlanned
Haitian Creole · 13M+ speakers
Spoken by nearly everyone in Haiti, while French dominates formal writing, so little text reflects how people talk.
Join the waitlist → - Runa SimiPlanned
Quechua · 7M+ speakers
A family of mostly oral varieties with little standardized text to learn from.
Join the waitlist →
Early access
Plug your model into how people really talk.
Tell us which languages your users speak and what your model gets wrong. We're opening access to a small group of teams first.