Most AI models were built for English first and learned Indian languages on the side. India needs language models that start from how India speaks, writes and works.
Most global models learn Indian languages from a thin slice of their data, and it shows. Param is pretrained from scratch on Indian data, with a tokeniser built for Indian scripts.
Government schemes, Indian names, festivals, exams, and sentences that mix Hindi and English. Context like this decides whether an answer is any use here.
Banks, hospitals and government departments often can't send data to a model hosted abroad. Param's weights are open, so it runs on your own servers in India.
Param2 uses only 2.4 billion of its 17 billion parameters for each word, so it costs far less to serve millions of people every day.
Our flagship text model, launched at the India AI Impact Summit in February 2026.
A mixture-of-experts language model that thinks step by step before it answers. It supports reasoning, code, maths and tool use across Indian languages.
Param2 is a mixture of experts. For each word it writes, a router hands the work to a few experts, so only about one parameter in seven is used.
Illustration. Each square is a slice of the model; the lit ones are working on the highlighted word.
Assamese, Bengali, Bodo, Dogri, Gujarati, Hindi, Kannada, Kashmiri, Konkani, Maithili, Malayalam, Manipuri, Marathi, Nepali, Odia, Punjabi, Sanskrit, Santali, Sindhi, Tamil, Telugu and Urdu, plus English.
Three real answers from Param2: maths, code with comments in Telugu, and a question about culture asked in Hindi.
A class has 30 students. The average score of the whole class in a test is 70. The average score of the top 10 students is 85. The average score of the bottom 10 students is 50. What is the average score of the middle 10 students?
Maths · Step-by-step reasoningOkay, let's see. The problem is about a class of 30 students. The average score of the whole class is 70. The top 10 students have an average of 85, and the bottom 10 have an average of 50. We need to find the average score of the middle 10 students.
First, I need to recall that the average score is the total sum of all scores divided by the number of students. So, for the entire class, the total sum of scores would be 30 students multiplied by 70, which is 2100. That makes sense because average is sum divided by number.
Wait, let me check my math again to make sure. Total class sum: 30 * 70 = 2100. Correct. Top 10 sum: 10 * 85 = 850. Bottom 10 sum: 10 * 50 = 500. Adding those: 850 + 500 = 1350. Subtracting that from the total: 2100 - 1350 = 750. Divided by 10 students: 750 / 10 = 75. Yep, that seems right.
The class has 30 students with an average score of 70. The total sum of all scores is 30 × 70 = 2100.
The top 10 students have an average score of 85, giving a total of 10 × 85 = 850. The bottom 10 students have an average score of 50, giving a total of 10 × 50 = 500.
Subtracting the top and bottom sums from the total gives the middle 10 students' total: 2100 − 850 − 500 = 750.
Dividing this by 10 students gives the average score of the middle 10 students: 750 ÷ 10 = 75.
The average score of the middle 10 students is 75.
Real output, not edited. Source: Param2 model card ↗
Write a binary search program in Python. All comments in the code must be written in Telugu language.
Code · TeluguUser binary search కోసం Python program కోరారు. Comments అన్ని Telugu లో ఉండాలి. Binary search సాధారణ algorithm, array, low, high, mid. Telugu లో వివరణలు ఇవ్వాలి. Code లో comments Telugu లో ఉండాలి, code logic correct ఉండాలి. Python లో function name in English, comments Telugu.
def binary_search(arr, target):
# ఈ ఫంక్షన్ లక్ష్య విలువను కనుగుకోవడానికి binary search ను అమలు చేస్తుంది
low = 0
high = len(arr) - 1
while low <= high:
mid = (low + high) // 2
# మధ్య విలువను పరిశీలిస్తాము
if arr[mid] == target:
# లక్ష్య విలువను కనుగొన్నాం, దాని index ను తిరిగి ఇస్తాం
return mid
elif arr[mid] < target:
# లక్ష్య విలువ పెద్దది, కాబట్టి low ను పెంచుతాం
low = mid + 1
else:
# target చిన్నది, కాబట్టి high ను తగ్గిస్తాం
high = mid - 1
# లక్ష్య విలువ లేదు, -1 ను తిరిగి ఇస్తాం
return -1
# ఉదాహరణ వినియోగం
arr = [2, 3, 5, 7, 11, 13]
target = 7
result = binary_search(arr, target)
print(f"లక్ష్య విలువ {target} యొక్క index: {result}")Real output, not edited. Source: Param2 model card ↗
भारत अपनी विविध संस्कृति के लिए जाना जाता है। भारत में अलग-अलग राज्यों की भाषाएँ, भोजन और त्योहार क्यों अलग होते हैं? इस सांस्कृतिक विविधता का भारतीय समाज पर क्या सकारात्मक प्रभाव पड़ता है?
In English: India is known for its diverse culture. Why do languages, food and festivals differ across India's states? What positive effect does this cultural diversity have on Indian society?
Culture · Hindiभारत अपनी विविध संस्कृति के लिए जाना जाता है, इसलिए प्रत्येक राज्य में अलग- अलग भाषाएँ, भोजन और त्योहार होते हैं।
भारतीय समाज पर सांस्कृतिक विविधता का सकारात्मक प्रभाव
इस प्रकार, भारत की सांस्कृतिक विविधता एक सकारात्मक प्रभाव है, जो एकता के साथ समृद्धि और पहचान को विकसित करता है।
Real output, not edited. Source: Param2 model card ↗
Our first family of models, small enough to run on modest hardware. Most are bilingual in Hindi and English; Param1 7B MoE covers 16 languages.
A 7 billion parameter mixture-of-experts model with 64 experts, pretrained from scratch on 4 trillion tokens. Built as a base model for fine-tuning.
A 5 billion parameter model for Hindi and English, with stronger maths and code than Param1 2.9B.
Param1 tuned to follow instructions and hold a conversation. The base for our domain models.
Our first foundation AI model: 2.9 billion parameters, trained from scratch on Hindi and English.
Param1 2.9B Instruct, trained further on specialist knowledge in Hindi and English.
For Ayurveda, trained on classical texts and clinical guidance.
For farming questions, from crops and soil to schemes.
For Indian finance, banking and accounting knowledge.
For Indian law, including question answering and summarisation.
Our benchmarks for these domains are on the datasets page.
The weights are open: download Param from Hugging Face or AIKosh and run it on your own servers, or talk to us about fine-tuning it for your work.