India’s homegrown AI model - Sarvam AI

Rajashree Rajadhyax
6 days ago
13 min read

TL;DR

Sarvam AI recently launched new India-focused AI models aimed at supporting multiple Indian languages.
I tested them using real-world prompts in English, Hindi, Marathi, and mixed-language queries.
The models show good potential in multilingual interaction and accessibility.
However, areas like factual accuracy, policy details, and deeper cultural nuance still need improvement.
Building AI for India is uniquely challenging due to language diversity, limited datasets, and complex policy information.
Despite the challenges, this is an important early step toward building AI systems designed for India.

Background

When Sarvam AI was launched, it felt like a proud moment for many of us in India’s technology ecosystem. For the past few years, most of the large language models we interact with have been developed in other parts of the world, primarily the US, China and others. These models are powerful and widely used, but they are built with a global audience in mind.

The launch of Sarvam AI was something different: the possibility of building foundational AI technology from India, for India. In Feb 2026, Bengaluru-based startup Sarvam AI released two large open-weight AI models, Sarvam-30B and Sarvam-105B, at the India AI Impact Summit in New Delhi. What made the announcement notable was that the company made the models publicly available under an Apache 2.0 licence, allowing developers and companies to experiment with them and even build commercial applications. These days I rarely get the time to sit and code, but this was too tempting to ignore. As someone building in the AI space myself, I was interested in seeing how such a system performs on real, everyday Indian queries, not benchmarks, but the kinds of questions people actually ask.

So I quickly set up the environment and decided to try out the APIs myself. In this article, I’m sharing some observations from that exercise, what worked well, where there were gaps, and what this tells us about the broader challenge of building AI for India.

Why India Needs Its Own AI Ecosystem

India has always adopted new technology very quickly. Whether it is smartphones, digital payments, or online services, people across the country have embraced technology at a very fast pace. While English is widely used, India is also a country with hundreds of languages. In everyday life, people often speak in their regional language or mix English with local languages in the same sentence. Take for example WhatsApp. Indians very commonly use Hinglish or use Roman script for native languages or some such combination for conversing. Because of this, AI systems that are built mainly for English-speaking users may not always work well for everyone in India. Building AI models in India gives us a better chance to create systems that understand our languages, our way of speaking, and the local context of things like government schemes, small businesses, and daily life.

What is Sarvam AI?

Most people today are familiar with tools like ChatGPT and the idea of large language models (LLMs) that can answer questions, write content, and have conversations. Sarvam AI is an Indian startup working in this same space. Its goal is to build AI models and infrastructure from India, with a strong focus on Indian languages and local use cases. The larger goal behind this effort is often called sovereign AI, which is about building core AI technology within the country instead of depending entirely on models developed elsewhere. By working on multilingual AI and India-specific applications, Sarvam AI aims to contribute to building an AI ecosystem that is created in India and better suited to Indian users. The company was founded in 2023 by Vivek Raghavan, Pratyush Kumar, and Ankit Jain.

Models Available

At the India AI Impact Summit 2026, Sarvam AI launched two large language models designed for Indian languages, Sarvam 30B, a 30-billion parameter model and Sarvam 105B, a 105-billion parameter model. These models are intended for general AI tasks such as conversation, question answering, and content generation. Alongside these, the company has also developed several specialised models. Bulbul is a text-to-speech model that generates natural-sounding voices in multiple Indian languages, while Saaras focuses on speech recognition. Sarvam has also introduced models such as Sarvam Vision, which can understand documents and images, and Sarvam Translate / Mayura, which are designed for translation across Indian languages. Together, these models form a broader AI stack aimed at supporting multilingual applications for businesses, developers, and public services in India.

A Note on Scale and Context

Before comparing Indian models with global LLMs, it is worth briefly looking at the difference in scale at which these systems are being built.

Global AI leaders such as OpenAI operate with extremely large levels of funding and computing infrastructure. As of early 2026, OpenAI has raised well over $100 billion and has access to massive GPU clusters through partnerships with companies like Microsoft and Amazon.

In contrast, Indian startups like Sarvam AI are working with far smaller resources. Sarvam has raised tens of millions of dollars and relies in part on the shared national compute infrastructure being built under the IndiaAI Mission. The gap in available compute is also significant. While global frontier labs operate GPU clusters running into hundreds of thousands of units, Indian AI companies are currently working with clusters in the thousands.

This difference in scale does not diminish the importance of building domestic AI capabilities. Instead, it highlights the ambition of the effort in building models tailored for Indian languages and use cases while operating with far more limited resources.

What the industry experts say

Several AI researchers and practitioners have also noted some interesting design choices in the Sarvam models. AI researcher Sebastian Raschka highlighted that the larger 105B model performs competitively with other models in its class such as GPT-OSS 120B and Qwen3-Next 80B across multiple benchmarks. He also pointed out that while Sarvam may not be the strongest coding model in benchmarks like SWE-Bench Verified, it performs surprisingly well on agentic reasoning and task completion tasks.

Another interesting observation comes from AI engineer Sebastian Sigl, who noted that Sarvam’s decision to use Grouped Query Attention (GQA) in the 30B model and Multi-Head Latent Attention (MLA) in the 105B model suggests the team made deployment-aware architecture decisions rather than simply scaling up the same design. He also pointed out that Sarvam’s custom tokenizer for Indian languages could be a significant advantage, since tokenizer efficiency directly affects inference cost, sequence length, and model performance on regional language text.

In simple terms, these design choices could make the models both more efficient to run and better suited for real-world applications in India.

The Use Cases I Tested

To get a practical sense of how the system performs, I decided to test Sarvam AI using a few everyday scenarios that are common in India. Instead of technical benchmarks, I focused on the kinds of questions real users might ask.

The prompts included a mix of languages and contexts:

A small business GST query in mixed Hindi and English
Questions about a government scheme in Marathi
A policy-related question in Hindi
General factual questions in Hindi, Marathi, and English
Some cultural and everyday topics in Marathi

The goal was not to run a formal benchmark, but to see how the system behaves when faced with multilingual, real-world queries that reflect how people in India actually communicate with AI systems.

Evaluation & Analysis

Based on these interactions, I evaluated the responses across a few key aspects such as language fluency, factual accuracy, and cultural understanding.

6.1 Language Fluency

One of the first things that stood out was the model’s ability to respond in multiple Indian languages. It handled Hindi, Marathi, and mixed-language inputs reasonably well and generated responses that were grammatically understandable.

The model was also comfortable switching between languages when the prompt itself contained mixed Hindi and English. This is important for India, where code-mixing is very common in everyday conversations.

Overall, the system showed good surface-level language fluency, especially for conversational responses.

Fig1: Snapshot of my conversation with Sarvam AI - code-mixed query

6.2 Cultural Depth

When the questions moved into culturally specific topics, the results were more mixed.

For example, when asked about Ukdiche Modak, a traditional Maharashtrian sweet, the model attempted to explain the dish and even provided a recipe. However, the response contained several inaccuracies. It incorrectly mentioned ingredients such as kokum jaggery and suggested a preparation method that involved frying the modaks, which is not how this dish is traditionally made. The explanation also included confusing phrases such as “कोरडे कोळी” when referring to ingredients, which clearly did not make sense in the context of cooking.

This example highlights an important challenge. India’s cultural landscape is extremely diverse, with regional foods, traditions, and terminology that are deeply specific to local communities. Capturing these nuances requires very strong cultural and linguistic training data. While the model was able to generate a response in Marathi, the inaccuracies suggest that deeper cultural grounding may still need to improve for such topics.

For AI systems that aim to serve users across India, understanding these kinds of cultural details will be just as important as understanding the language itself.

Fig2: Snapshot of my conversation with Sarvam AI - query to assess cultural depth

6.3 Factual Accuracy

For general knowledge questions, the model was able to produce responses that were structured and confident. However, in some cases the information provided was not fully accurate.

For example, I asked whether a census was conducted in India in 2021. The response stated that the 2021 census had been conducted in phases but was delayed due to the COVID-19 pandemic. In reality, the 2021 Census was postponed and has not yet been conducted, which makes the response factually incorrect.

In another example, when asked about the number of states in India and literacy rates, the model correctly identified Kerala as having one of the highest literacy rates. However, the explanation mixed data points and comparisons between states and union territories in a way that could easily confuse readers.

These examples show that while the model can produce answers that sound well structured and authoritative, the underlying facts may not always be reliable. This highlights a common challenge with many language models today that they can produce very confident answers even when the underlying information may not be correct.

Fig3: Snapshot of my conversation with Sarvam AI - query to assess factual accuracy

6.4 Policy & Compliance Reliability

One of the more interesting tests involved a small manufacturing unit in Mumbai asking about GST registration requirements. The prompt asked:

“Main Mumbai mein ek small manufacturing unit chalaata hoon. Mujhe GST registration kab compulsory hota hai? Agar mera turnover 4 crore hai toh kya karna padega?”

The response provided a clear explanation of GST registration thresholds and described the compliance steps involved. It explained that registration becomes mandatory when turnover crosses certain limits and mentioned filing requirements such as GSTR-1 and GSTR-3B. From a readability perspective, the answer was structured and easy to understand. However, some of the specific details were not entirely accurate. For example, the explanation around composition scheme eligibility and turnover limits was not precise. The response also mentioned penalties and thresholds that did not fully match the actual GST rules. This highlights an important consideration. When AI systems are used for business, regulatory, or compliance-related queries, accuracy becomes especially critical. Even small numerical or regulatory mistakes could lead users to misunderstand their obligations.

While the model showed promise in explaining concepts in simple language, policy and compliance use cases will require a higher level of factual precision and reliable grounding before they can be trusted for real-world decision making.

6.5 Hallucination Behavior

Like many large language models, Sarvam AI sometimes generated answers that sounded plausible but were not fully grounded in verified facts. This was visible in a few areas, particularly when the system was asked about events or government schemes. Instead of indicating uncertainty, the model often produced a structured answer even when the information appeared to be speculative.

This behavior is not unique to Sarvam, it is a broader challenge across the LLM ecosystem, but it highlights the importance of strong grounding mechanisms and better uncertainty signaling.

6.6 Tone & Accessibility

One area where the model performed well was tone and readability. The responses were generally polite, structured, and easy to understand. For users who may not be comfortable with technical language, this style of response can make AI systems more accessible and less intimidating. The model also attempted to provide step-by-step explanations when appropriate, which is useful for informational queries. From a user-experience perspective, this conversational and structured tone is an encouraging sign for broader adoption.

Where It Shows Promise

Despite some of the gaps observed during testing, there are several areas where the model shows clear promise. One of the most encouraging aspects is its ability to operate across multiple Indian languages. The model was able to respond in Hindi and Marathi and generally maintained the script and conversational flow reasonably well. For a country like India, where a large number of users are more comfortable interacting in regional languages than in English, this is a very important capability. Another positive aspect is the tone and accessibility of the responses. In many cases, the model explained concepts in a simple and conversational way. Even when answering policy or factual questions, the responses were structured in a way that a non-technical user could understand. The model also showed potential in India-specific queries. Questions related to government schemes, business regulations, and everyday practical topics were handled with an attempt to provide locally relevant explanations. While the factual accuracy still needs strengthening in some cases, the direction is encouraging.

Perhaps the most important takeaway is that this is an early step in building AI systems trained with Indian languages, contexts, and use cases in mind. The ability to handle multilingual prompts and mixed language queries already shows that the foundation is being laid for something that could become significantly stronger with further iterations and training.

Challenges in building AI in India

Building AI for India is not an easy problem. The country is incredibly diverse in terms of language, culture, and the way people communicate. Because of this, creating AI that truly understands Indian users comes with several challenges.

Many Languages and DialectsIndia has hundreds of languages and dialects. Even within the same language, the way people speak can change from one region to another. Training a model to understand all these variations is a huge task.

Limited Digital Data in Regional LanguagesEnglish has a massive amount of content available online, which makes it easier to train AI models. Many Indian languages, however, have much less digital content. Important information—such as government policies, scheme details, or local administrative rules—is often published in regional languages and sometimes in formats that are not easily accessible for AI training. This makes it harder to build models that can reliably understand and respond to these topics.

Code-Mixed CommunicationIn everyday conversations, many Indians mix languages. For example, someone might start a sentence in Hindi and finish it in English, or mix Marathi and English. Teaching AI to understand this natural style of communication is technically challenging.

Cultural NuancesIndia’s culture is very rich and very local. Food, traditions, festivals, and everyday references can be very different from one state to another. For AI to understand these details well, it needs very deep and diverse training data.

Building the EcosystemTraining large AI models requires huge computing power, infrastructure, and investment. Countries like the United States have had a long head start in building this ecosystem. For Indian startups, the challenge is not just building the model, but also building the supporting ecosystem around it.

When you look at all these factors together, it becomes clear that building AI for India is a complex and ambitious effort.

What Needs Strengthening

While the early results are encouraging, there are a few areas where the model could become significantly stronger. These improvements would make the system more reliable for real-world use in India.

Stronger Grounding in Current Affairs For questions related to recent events or evolving situations, the model would benefit from stronger grounding in up-to-date information. News, policy decisions, and administrative changes happen frequently in India, so keeping the model aligned with current information will be important.
Better Integration with Verified Government Sources Many users will naturally turn to AI to ask about government schemes, policies, and benefits. Integrating information from verified government sources could help improve the reliability of these answers and reduce the chances of incorrect or outdated details being presented.
Clearer Signaling When the Model Is Uncertain In some cases, the model responds with confidence even when the information may not be fully accurate. It would be helpful if the system could signal uncertainty more clearly—for example by indicating when it is not fully sure or when information should be verified from official sources.
Greater Precision in Numeric and Compliance Information For topics such as taxation, regulatory thresholds, or eligibility criteria for schemes, even small numerical inaccuracies can create confusion. Strengthening precision in these areas would make the model more useful for practical business and compliance-related questions.
Deeper Understanding of Regional Languages and Context While the model can already respond in languages like Marathi and Hindi, deeper understanding of regional vocabulary, cultural references, and everyday expressions would make the responses feel more natural and accurate. India’s linguistic diversity makes this an important area for continued improvement.

How You Can Try Sarvam AI Today

If you are curious to explore Sarvam AI yourself, there are a few ways to try it depending on whether you are a general user, someone who wants to experiment with prompts, or a developer looking to build applications.

Indus – Chat Interface (Similar to ChatGPT)

The simplest way to experience Sarvam’s models is through Indus, a chat-based interface developed by Sarvam AI. The experience is similar to interacting with ChatGPT where you can type questions in a conversational format and receive responses instantly. Indus is designed with Indian users in mind. It supports multiple Indian languages and can handle prompts written in Hindi, Marathi, or mixed language sentences where English is blended with an Indian language. This makes it particularly useful for testing how the model performs in everyday Indian communication styles. Users can try asking general knowledge questions, local topics, or even region-specific queries to see how the model responds. Try it here: https://indus.sarvam.ai

Sarvam AI Playground (For Testing Models)

For users who want to experiment more deliberately with prompts, Sarvam also provides an online Playground environment. This interface allows you to directly test the AI models and observe how they respond to different types of inputs.

The playground is useful if you want to run structured experiments, for example comparing responses across different languages, testing factual questions, or trying code-mixed prompts. It is a convenient way to explore the behavior of the models without needing to write any code. Researchers, testers, and curious users can use this environment to better understand how the models perform across different scenarios.

Try the playground here: https://try.sarvam.ai

Developer Platform and APIs

For developers who want to build applications using Sarvam AI, the company provides a developer platform with API access. This allows the models to be integrated into products such as chatbots, enterprise assistants, voice interfaces, and multilingual customer support systems. Through the APIs, developers can access capabilities such as conversational AI, translation, speech-to-text, and text-to-speech. This opens up possibilities for building solutions that are tailored for Indian users and multilingual environments. The documentation and technical guides provide details on how to authenticate, send prompts to the models, and integrate them into applications. Developer documentation: https://docs.sarvam.ai

Closing Thoughts

Building large language models is one of the hardest technology problems to solve today. When you look at it from that lens, the effort to build AI systems that understand India’s languages, policies, and cultural nuances is both ambitious and important.

From my own testing, I could see that there are areas where the models still need to improve. But at the same time, it is clear that the foundations are being laid for something meaningful, AI systems that are being built with Indian users, Indian languages, and Indian realities in mind.

As someone working in the AI space myself, I also appreciate how difficult this journey is. Building AI at this scale requires enormous amounts of data, compute, research, and persistence. Doing this while trying to represent the diversity of India makes the challenge even bigger.

What we are seeing today may just be the beginning of India’s AI journey. If the ecosystem continues to invest in better datasets, stronger infrastructure, and deeper language understanding, the future could be very exciting.

For now, the best thing we can do is try these models ourselves, experiment with them in our own languages, and support the efforts being made to build AI from India, for India.