Indian artificial intelligence startup Sarvam has unveiled two innovative large language models (LLMs) specifically designed for Indian languages. The announcement, made on Wednesday, introduces a 30 billion parameter model and a more extensive 105 billion parameter variant, both optimized for real-time applications.
The 30 billion parameter model is particularly notable for its ability to manage real-time conversations with a 32,000 context window, significantly lowering inference costs. Meanwhile, the larger model, equipped with an impressive 128,000 token capacity, is tailored for more intricate tasks, enhancing its functionality for complex problem-solving.
At the launch event, Sarvam shared multiple milestones demonstrating the new models’ competitive performance against established international benchmarks, including Gemma 27B, Mistral-32-24B, and GPT-OSS-20B. “It is on par with most other open and closed frontier models of its class, designed to excel in complex reasoning tasks,” stated Sarvam co-founder Pratyush Kumar.
Kumar further emphasized that the 105 billion parameter model competes effectively with models like DeepSeek R1, which boasts 600 billion parameters. “Trained from scratch at one-sixth the size, this model provides intelligence comparable to its predecessors and is more cost-effective than Google’s Gemini Flash, outperforming many of its benchmarks,” he added.
A highlight of the launch was Vikram, Sarvam’s AI chatbot, demonstrating its multilingual capabilities by conversing in several Indian languages, including Hindi and Punjabi. The name “Vikram” pays homage to Vikram Sarabai, a celebrated Indian physicist and astronomer.
Sarvam was selected by the India AI Mission to establish the nation’s first sovereign LLM ecosystem, with plans for an open-source model comprising 120 billion parameters. This initiative aims to enhance governance and public service access through applications such as 2047: Citizen Connect and AI4Pragati.
In tandem with Sarvam’s efforts, Soket is also working on India’s first open-source 120 billion parameter foundation model, which will cater to the country’s linguistic diversity, targeting sectors like defense, healthcare, and education. Meanwhile, other startups like Gnani and Gan AI are developing their own models focused on multilingual capabilities, with Gan AI targeting a 70 billion parameter model optimized for text-to-speech functions.
Both Sarvam’s models support all 22 official Indian languages and are designed for voice-first interaction, addressing the growing importance of voice AI in making technology accessible to the wider population.
This launch arrives against the backdrop of OpenAI’s introduction of IndQA, a benchmark aimed at evaluating AI models’ comprehension of questions in various Indian languages. OpenAI is focusing on India, which ranks second globally in its user base after the United States. Similarly, Anthropic has integrated 10 Indic languages into its Claude AI, including Hindi, Bengali, and Urdu.
With approximately $50 million in funding from notable investors such as Khosla Ventures, Lightspeed Venture, and Peak XV, Sarvam is poised to make significant strides in the AI landscape, setting the stage for advancements that will benefit diverse linguistic communities across India.
Original Source: https://www.business-standard.com/technology/tech-news/ai-startup-sarvam-launches-two-made-in-india-large-language-models-126021801397_1.html
Category :
Tags:
Publish Date: 2026-02-18 20:47:00
