Sarvam's new AI models target the masses with multilingual support
What's the story
Sarvam AI has launched two new artificial intelligence (AI) models, the Sarvam 30B and Sarvam 105B, at the ongoing AI Impact Summit 2026. The key highlight of both these models is their ability to support Indian languages such as Punjabi, Marathi, and Hindi. The company's co-founder Pratyush Kumar emphasized that these models have been trained with a focus on affordability for the masses.
Model overview
Sarvam 30B model
The Sarvam 30B model is geared toward a wide range of applications such as multilingual text generation, conversational AI, translation, and content creation. With 30 billion parameters, it strikes a balance between efficiency and power. Kumar demonstrated several benchmarks at the launch event showing that Sarvam 30B outperforms some well-known AI models in the market.
Practical use
Chatbot Vikram can converse in multiple Indian languages
During the summit, Sarvam AI showcased a practical application of its 30B model. The demonstration showed how even a feature phone could be used to converse with an AI, thanks to this model. The company's chatbot, Vikram, can communicate in several Indian languages including Punjabi and Hindi. The name "Vikram" was chosen in honor of Vikram Sarabhai, an eminent Indian physicist and astronomer.
Advanced capabilities
Sarvam 105B model
The Sarvam 105B model is a more powerful version, designed to handle complex reasoning and high-level language tasks. With a context window of 128K tokens, it can process and retain much larger amounts of information than most similar systems. This enables it to handle sophisticated use cases such as long-form summarization, in-depth analytical queries, and nuanced dialog generation.
Global impact
Can the new models challenge global giants?
Sarvam AI claims that the scale and efficiency of its 105B model allow it to outperform global competitors, including the DeepSeek's R1 model. This marks a major milestone for India's AI capabilities on the world stage. During the summit, Sarvam demonstrated this model's real-world potential by uploading a company's balance sheet and having it answer complex financial questions with precision in seconds.