A monsoon is not one storm. It is a season-long conversation between warm oceans, winds, mountains and land. A small change in one place can shift rain hundreds of kilometres away, so a forecast must track the whole system—not just tomorrow's clouds.
Indian-language AI.
Many ways to think.
Ask, write and explore in Hindi, Gujarati and other Indian languages. Aditya brings language research and practical work together.
Currently testing with our internal team. For access enquiries, email us.
Your language.
Your way into the idea.
A question can begin in your own script, in Latin letters, or with words from more than one language. Explore examples in Hindi, Gujarati, Marathi, Bengali, Tamil and Telugu, alongside English.
Our Indian-language evaluations include native and romanised input. The language pages show illustrative prompts and explain the measured scope, so you can distinguish a way of asking from a result we have actually tested.
Evaluation coverage is not a guarantee for every phrasing. Read the examples and measured scope.
See a way of working.
Start with a question.
Choose a writing illustration. These are editorial examples, not live model answers or measured results.
Clear thinking
Illustrative writing. Review important answers before relying on them.
Example promptExplain why monsoon forecasts are difficult, without jargon.
No model runs on this page. Email the lab about access.
Possibilities
Conversation with shape
Move from a half-formed question to a clear decision without losing the thread.
Begin with the context you have: a rough note, an unfamiliar topic or a decision with competing constraints. Our writing examples show how a response might clarify the question, explain the assumptions and suggest what to check next.
- Name the real question
- Bring context into view
- Move toward a next step
Illustrative workflow, not a live model demonstration.
Explore the model
A NAME TO BEGIN WITHआदित्य
One intelligence. Many ways to think.
Research you can read.
Limits you can see.
- What was tested
- What it was compared with
- What remains open
Our evaluation report covers Indian-language maths, knowledge, refusal and over-refusal. Each comparison names the sarvam-m base model and the evaluation used, rather than combining different tasks into one headline score.
GSM8K-Indic checks numerical problem solving; MMLU-Indic checks knowledge questions. The safety tests ask a different question: when does the model refuse a harmful request, and when does it turn away a harmless one? Read those outcomes together to understand the trade-off.
Saved scores describe those tests, not every conversation. They do not establish universal reliability, safety or superiority over another assistant. Our research pages keep measured results separate from architecture ideas and workflows that still need testing.
Capability should be useful.
Trust should be earned.
A useful model should make its work understandable, not ask people to trust a polished answer on its own. When a task has real consequences, the boundaries and the choice to proceed should remain visible. Its answers should be grounded in evidence, and its limitations stated plainly. That is how capability becomes something people can use with confidence.
A useful handoff should say what was checked, which sources support the answer and what remains uncertain. Our coding-agent notes examine task records, approvals and verification. The evaluation report provides the narrower evidence for model behavior; neither a polished interface nor an architecture diagram is proof that a task will always succeed.
Open the research programBuild with the Aditya API
Use an OpenAI-compatible API for Indian-language AI in your own application. Aditya is being tested by our internal team; email the lab about access before integrating.
A familiar interface.
A different starting point.
from adityalab import Aditya
client = Aditya() # reads ADITYA_API_KEY
reply = client.chat.completions.create(
model="aditya-latest",
messages=[{"role": "user",
"content": "Explain an API in Hindi."}],
conversation="new",
)
print(reply.text)Install: pip install adityalab
Example only · runs on your server. Never expose your API key in browser code.
Open the developer platform API guide
Official SDKs
Use our Python and JavaScript/TypeScript SDKs. The API remains OpenAI-compatible.
Evaluation
Indian-language AI
Explore native-script and romanised input. Read the measured scope.
Pricing
Prepaid, per token
API calls are charged from your balance. See prices and billing rules.
Billing guide
Trace each request
Every response returns an X-Request-Id for troubleshooting.
Pricing is per token in ₹. Pay as you go: ₹25 per million input tokens and ₹100 per million output tokens, charged from your prepaid balance.
View pricingBuilt with care.
Open about the work.
Aditya brings Indian-language research and practical, reviewable agent workflows together.
Aditya is the general model. Aditya Code is the coding agent. Our research pages separate implemented mechanisms, saved measurements and ideas that still need testing.
Explore the research programThe questions
behind the work.
Follow the ideas, design decisions and open questions shaping Aditya. These are notes from a build, not a list of promises.
What is Neural Intelligence Fabric?
A research direction for composing, checking and reusing capability—not a claim of established general intelligence.
Read the note
Why tool calling is not enough
The gap between proposing a function call and completing useful engineering work.
Read the note
Building coding agents that can prove the work
Task records, checks and uncertainty belong alongside the answer.
Read the noteBring a real question.
Meet a broader intelligence.
Testing is currently limited to our internal team. Email the lab to discuss your use case and access.
आपका अगला सवाल क्या है?
Bring a question from your own work. We are looking for the useful, difficult, human shaped tasks that should guide this research.
Contact Aditya Lab Internal testingFrequently asked questions
What is Aditya AI?
Aditya AI is an Indian-language AI research model and assistant from Aditya Lab. The website's chat and workspace examples are illustrations, not live model conversations.
Which languages does Aditya AI support?
Our evaluation covers Hindi, Gujarati, Marathi, Bengali, Tamil, Telugu and English. That measured scope does not guarantee the same quality for every request or language.
Does it understand romanised Hindi or Gujarati, including Hinglish and Gujlish?
The Indic evaluation includes Hindi and Gujarati written in Latin letters as well as native script. These saved tests do not establish that every Hinglish or Gujlish phrasing will work reliably.
How can I get access to Aditya AI?
Aditya is being tested by our internal team. Email [email protected] to discuss your use case and access; no general-access date has been announced. API-key calls are paid per token from a prepaid balance.
How is Aditya AI different from ChatGPT or Sarvam?
Aditya Lab focuses on Indian-language behavior and visible, reviewable agent work. Our published comparisons use sarvam-m as a base model on named evaluation suites, not as a general ranking against Sarvam or ChatGPT.
What do the research pages show?
The research pages distinguish saved measurements from architecture ideas and illustrative interfaces. Reports include methods and limitations so readers can interpret the results.
Is my data used for training?
Your conversations are not used to train or improve our models unless you opt in through your account settings; this is off by default. You can turn it off at any time, after which new conversations are not used. See the privacy policy for retention and connected-service details, and do not email sensitive material.