Sarvam 30B-A1B & 105B-A9b launched at India AI Impact Summit 2026

by

New Delhi: India AI Impact Summit 2026 brought fresh news from the AI world on Wednesday. On the third day of the summit in New Delhi, Sarvam AI rolled out two new models: Sarvam 30B and Sarvam 105B. These caught attention because they handle Indian languages such as Punjabi, Marathi, Hindi and several others quite well. Co-founder Pratyush Kumar stepped up during the event and demonstrated how the company trains these models to serve the population at a reasonable price. It shows India’s push to build its own AI tools that fit local needs instead of depending fully on foreign ones.

The summit runs from February 16 to 21 under the IndiaAI Mission. Experts, officials and companies gather to talk about AI’s role in growth and daily life. Sarvam’s launch fits into those discussions. People at the event saw live demos. Kumar highlighted benchmark results. Many walked away thinking about how these could reach regular users across the country.

Sarvam 30B focuses on wide, everyday tasks

Sarvam 30B has 30 billion parameters. Engineers built it to balance power with efficiency. It works for things like generating text in multiple languages, chat conversations, translation and making content. The model suits spots where quick responses and low costs count a lot. In the launch demo, Kumar showed several benchmarks. He pointed out that Sarvam 30B has already defeated a few famous AI models. That got folks talking, especially since it performs strong in tests that matter for real use.

On stage demos show it runs well for real-time apps. Some even say it could bring AI assistants to basic feature phones. Imagine a shopkeeper in a small town chatting in Hindi or Punjabi without needing fancy hardware. That kind of reach excites people here.

Sarvam 105B takes on harder jobs

Sarvam 105B stands as the stronger one. It handles complex reasoning and tough language tasks. The model keeps a context window of 128K tokens. This lets it deal with big chunks of info, like long summaries, detailed questions or careful talks. According to Sarvam AI, the model’s scale and efficiency allow it to outperform global peers, including the DeepSeek R1 model. That marks a significant leap for India’s AI capabilities on the international stage.

Both models pay close attention to Indian languages. Global tools often trip up on our accents or mixed words. Sarvam tries to fix that. Kumar explained in the demo how they train the models to serve the population at a reasonable price.