Sarvam AI, an Indian startup, is addressing India’s AI adoption challenges stemming from its diverse linguistic landscape. With a focus on developing large language models (LLMs) tailored for India’s 22 languages and multiple scripts, Sarvam aims to bridge the gap between technical innovation and enterprise implementation.
Sarvam’s approach includes two indigenous models – a 30B and 105B open-source models, utilizing a unique mixture-of-experts system. Notably, Sarvam’s custom tokenizer optimized for complex Indian dialects like Odia and Manipuri reduces computational requirements, making it more efficient.
Sarvam’s models have been deployed in real-world applications, such as a voice-first rural banking assistant developed by Aphelion Labs. Industry experts recognize Sarvam’s strength in Indic language processing, especially in non-Hindi and non-English contexts.
However, Sarvam faces challenges, including the lack of local execution formats and support for inference frameworks, posing obstacles to widespread adoption. The company is actively engaging with the Indian startup ecosystem through initiatives like offering credits and founder sessions to address these challenges.
Source: Inc42 Media