India’s Homagron AI Startup Sarvam has launched its latest language model, Sarvam-M, which is making waves in the technical community for both good and not good reasons. The model is being praised for focusing on Indian languages, mathematics and programming works, but is also facing criticism for not being “good”. The drama around the AI ​​company has created even more interest. If you also have questions, then what is the Saravam-M model here, it has a breakdown, why it matters, and the AI ​​company is facing backlash.

What is really Sarvam-M?

Sarvam-M is a big language model, or LLM, developed by Indian Startup Saravam AI. These types of models are trained to understand and generate human-like text, and they power tools such as chatbots, translation software and educational apps. Saravam-M is based on a small model called Mistral Small and is expanded in a very large system with 24 billion parameters, basically knobs and dials that help it to process the process and learn from data.

In simple words, Sarvam-M is like a very smart AI assistant that can handle a wide range of tasks, answering complex mathematics questions to understanding and answering in Indian languages ​​like Hindi, Bengali, Gujarati. It is different that it is built with India, supports 10 local languages ​​and offers a strong performance in both language and argument related tasks.

How was it made?

Sarvam-M was trained using a three-step process:

Supervised Fine-Tuning (SFT): The model was included in this stage to help learn high-quality questions and answers. The team ensured that the reactions were relevant, less biased and culturally suitable. This helped the model be good in both everyday interaction and more complex problems.

Learning reinforcement with verification awards (RLVR): In this phase, Saravam-M was further improved using data related to instructions, programming and mathematics. It was taught better to follow the instructions and to think more logically using feedback loops and careful designed tasks.

Estimate adaptation: This final stage included the model to run rapidly and more efficiently. FP8 quantation (a way to simplify data without losing accuracy) and better decoding methods helped improve model speed and performance, although there were still some issues with handling high traffic.

What can Saravam-M do?

The model is designed to give strength to various real -world applications. This can be used for:

Convious AI, Which will basically mean that it can power the chatbots and virtual assistants

machine translationWhere it can be used to translate between English and Indian languages

Education, Keeping in mind the ability to solve mathematics problems, to help students prepare for competitive exams like JEE. In fact, one of the members of the Saravam team shared the results that Sarvam-M’s “Think” mode gave the correct answer to many JEE advanced level questions in Hindi, which could be a big step for Indian students to make such equipment useful.

How does it compare with other models?

Saravam-M has shown impressive results in some areas. In a test, which introduced mathematics with romantic Indian languages, the model gained more than 86 percent improvement, beating some other famous models. It performed better than Meta’s Lama -4 scouts at several benchmarks and was equal to very large models like Lama -370B and Google’s Jemma 3 27B.

However, it underwent a slight underperform in English knowledge tests, with about 1 percent less accuracy than others. Nevertheless, the model stands for its Indian language skills and logic abilities.

So, why backlash?

Despite all the technical achievements, the model was not warmly welcomed. On the face of a hug, a platform where developers can download and test the AI ​​model, Saravam-M was downloaded only 334 times in the first two days. Some critics saw it as a sign of failure.

Dedi Das, an investor from Menlon Ventures, called the response “embarrassing”, saying that there is little interest in such work. He compared it to a separate model created by two Korean college students, with around 200,000 downloads quickly.

This led to a debate. Sarvam-M supporters, including Sachdeva, referred to the company, defended the model, highlighting their benchmark results and adaptation process. He even posted evidence of the model performance on social media.

Another user, who works in AI4Bharat, said that the actual achievement was not just a model, but a method used to train it. He said that it sets a strong base for the construction of other Indian developers.

Meanwhile, Sarvam’s co-founder, Vivek Raghavan, called Sarvam-M “Stepping Stone” towards the construction of his AI system of India. The company is one of the few selected under the India Mission of the Government of India to develop a sovereign LLM for the country.

Zoho founder Sridhar Vambu also urged people not to focus on immediate success. He said that most of the products take time to find their place, and praised Sarvam for their efforts. “Keep fighting well,” he encouraged.