Bengaluru-based Sarvam AI has launched its Indus conversational app in beta, featuring a custom 105B parameter model optimized for Indian languages and contexts. The release targets India's multilingual population but faces scalability and accuracy challenges common to regional LLMs.
Sarvam AI, the Bengaluru-based startup developing foundation models for Indian languages, has launched its Indus chat application in public beta. The web and mobile app is powered by the company's proprietary Sarvam 105B model, positioning itself as India's answer to global chatbots like ChatGPT.
The core technical claim centers on the Sarvam 105B model's optimization for India's linguistic diversity. Unlike models primarily trained on English corpora, Sarvam's architecture incorporates training data from 22 scheduled Indian languages including Hindi, Bengali, Telugu, Marathi, and Tamil. The model uses a custom tokenizer that handles script variations and code-switching patterns common in Indian multilingual communication.
According to Sarvam's technical documentation, the 105B parameter model achieves 15-20% better perplexity scores on Indian language benchmarks compared to equivalently sized multilingual models. Early testing shows improved handling of culturally specific queries – from explaining agricultural concepts in regional dialects to parsing legal terminology from vernacular documents.
Practical applications demonstrated include:
- Real-time translation between Indian languages while preserving context
- Generating government scheme explanations in local dialects
- Assisting with vernacular content creation for education and media
However, significant limitations emerge upon scrutiny. The model currently supports only text interactions, lacking the multimodal capabilities of competitors. Token efficiency remains problematic for longer-form Indian languages like Malayalam, where context windows deplete 40% faster than in English conversations according to independent tests.
Infrastructure challenges loom large. Serving a 105B parameter model requires substantial compute resources – Sarvam's current implementation reportedly experiences latency spikes during peak usage. The company hasn't disclosed its scaling strategy for supporting millions of potential users across India's varied network conditions.
Critically, benchmark comparisons against alternatives like Meta's Llama 3 or Mistral's models reveal performance gaps in complex reasoning tasks. When tested on the Indian Legal Judgment Prediction dataset, Sarvam 105B trailed Llama 3-70B by 12 percentage points in accuracy.
The launch comes amid India's push for sovereign AI capabilities. Sarvam recently secured $41 million in funding from investors including Lightspeed Venture Partners and Khosla Ventures. The Indus app represents India's most ambitious attempt at a homegrown LLM product, though real-world performance beyond controlled demos remains unproven.
Technical resources:
As beta testing expands, key questions persist about inference costs, hallucination rates in low-resource languages, and the feasibility of maintaining competitive model performance against globally scaled alternatives. For now, Indus serves as a significant milestone in regional AI development, though its long-term viability hinges on solving scalability challenges unique to India's linguistic landscape.
Comments
Please log in or register to join the discussion