Sarvam AI Sarvam-1 AI Language Model

There is another AI language model released by the emerging Indian generative AI player, Sarvam AI—Sarvam-1. This open-source paradigm has been developed exclusively for the Indian language milieu. Besides English, there are provisions to select 10 South East Asian and Indian languages, like Bengali, Hindi, Tamil, etc. Like before, the models were Sarvam 2B, which was launched in August 2024, and later the Sarvam 1 in October 2024.

Sarvam-1: Small Language Model

• Sarvam-1: Small Language Model (SLM) for 2B params | Microsoft, Phi-3 Mini-3.8B fitted.
• Yotta services 1,024 GPUs, and its training works on the NVIDIA NeMo framework.
• Sarvam-2T: A training corpus is built to tackle the unavailability of quality training data for Indian languages.
• Sarvam-2T is a 2 trillion-token dataset that includes equal parts of all ten languages already supported.
• Sarvam-1 engineers Indic Varnmala scripts more efficiently than existing LLMs and need fewer tokens per word.
• Sarvam-1 set a new record on the TriviaQA benchmark with an accuracy of 86.11 for Indic languages, outdoing Meta’s Llama-3. 1 8B.
• Sarvam-1 is appropriate for real-world use cases, including deployment on edge devices that often have limited computational capacity.
• You can download Sarvam-1 from Hugging Face, an open-source hosting platform.