India's first open-source AI model debuts, supports 10 Indic languages
What's the story
Sarvam AI, a Bengaluru-based artificial intelligence (AI) start-up, has introduced Sarvam 2B, India's first open-source foundational model.
The model is trained on an extensive dataset of four trillion tokens and can understand instructions in 10 Indic languages.
These languages include Hindi, Tamil, Telugu, Malayalam, Punjabi, Odia, Gujarati, Marathi, Kannada and Bengali.
Last year, the AI company secured $41 million in funding from investors such as Lightspeed Peak XV Partners and Khosla Ventures.
Model comparison
Sarvam 2B: A unique addition to small language models
Vivek Raghavan, co-founder of Sarvam AI, highlighted that Sarvam 2B belongs to a category of Small Language Models (SLMs).
This group also includes Microsoft's Phi series models, Llama 3 8B and Google's Gemma models.
Raghavan emphasized the uniqueness of their model stating, "This is the first open-source foundational model trained on an internal dataset of four trillion tokens by an Indian company with compute in India."
Model application
Sarvam 2B: A tool for Indic language tasks
The Sarvam 2B model will be accessible on Hugging Face and is particularly useful for Indic language tasks such as translation, summarization, and understanding colloquial statements.
The start-up has made the model open-source to encourage further research and development.
This move also supports the creation of applications based on it.
Audio innovation
Sarvam AI launches Shuka 1.0, an open-source audio language model
In addition to Sarvam 2B, the start-up also launched Shuka 1.0 at an event in Bengaluru today.
This is India's first open-source audio language model and serves as an extension of the Llama 3 8B model to support Indian language voice input and text output.
Raghavan explained that "the audio serves as the input to the LLM, with audio tokens being the key component here."
Model performance
Shuka 1.0 outperforms existing models in speed and accuracy
Shuka 1.0 is claimed to be six times faster than Whisper + Llama 3 and offers higher accuracy across the 10 languages compared to its counterparts.
The start-up's goal is to enhance this model to sound more human-like in the future.
This development signifies a significant advancement in AI technology catering specifically to Indian languages.
AI agents
Sarvam AI introduces voice-based, multilingual agents
Sarvam AI also unveiled Sarvam Agents, voice-based multilingual agents designed to address specific business challenges.
These agents can be integrated by contact centers or sales teams of various enterprises through three channels—telephony, WhatsApp, and within an app.
Raghavan explained the functionality of these agents stating "These agents can also be very contextual... The agent will be context-aware so it knows where you're asking from."