Sarvam Challenges Deepseek On Benchmarks

2026-03-11 · Source: AIM Network · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Novice, medium

Summary

Bengaluru-based Sarvam AI has released two open-weight large language models, Sarv 30B and Sarv 105B, trained in India using compute from the India AI mission. The 105B model shows strong performance in agentic browsing tasks, outperforming DeepSeek-Coder-1 on BrowseCamp, but trails rivals on coding benchmarks like SWBench. A key strength of the Sarvam models lies in their understanding of 22 Indian languages across 12 scripts, facilitated by a custom tokenizer. The 30B model is optimized for conversational AI, while the 105B model powers Sarvam's AI chatbot, "Indus." Community reception from Indian startup founders has been positive, with plans for integration into new products. Sarvam is also engaging with enterprise customers and partnering with Indian state governments, including Maharashtra and Odisha, and the State Bank of India, to drive adoption and contribute to India's open-source AI ecosystem.

Key takeaway

For NLP Engineers or CTOs evaluating LLMs for the Indian market, prioritize Sarvam AI's models for their deep understanding of 22 Indian languages and cultural nuances. While global benchmarks offer a rough idea, the models' real-world performance in Indic tasks and enterprise adoption potential, including government partnerships, suggest a strong foundation for localized applications. Consider integrating these models to serve Indian customers effectively and contribute to the national AI stack.

Key insights

Sarvam AI's models prioritize Indic language proficiency and real-world utility over global benchmark dominance.

Principles

• Localization drives AI adoption in diverse markets.
• Open-source contributions foster national AI ecosystems.

Method

Sarvam AI trained its models using Indian compute, collected data from diverse Indian sources (web, books, internet), and developed a tokenizer covering 22 Indian languages across 12 scripts to understand cultural nuances.

In practice

• Integrate Sarvam 30B for customer service conversations.
• Utilize Sarvam 105B for Indian language text processing.
• Explore Sarvam models for government digital public infrastructure.

Topics

• Large Language Models
• Open-weight Models
• Indian Languages
• AI Benchmarking
• Enterprise AI Adoption

Best for: NLP Engineer, CTO, VP of Engineering/Data, Tech Journalist, AI Product Manager, AI Engineer

Related on AIssential

• India’s LLM Revolution: The Rise of Indigenous AI for a Multilingual Nation — Indigenous LLMs are crucial for nations with high linguistic and cultural diversity, overcoming limitations of global models.
• NRITYAM: Language Models Meet Art and Heritage of Dance — NRITYAM is a benchmark evaluating language models' cultural comprehension in global dance traditions across 12 languages.
• CARTE: A Benchmark for Mapping Language Model Knowledge Across France — LLMs exhibit significant knowledge gaps in fine-grained, geographically specific regional understanding, particularly within countries.
• BharatGen is set to launch Param2 17B at the India AI Summit — India is pursuing "AI independence" through culturally relevant, cost-efficient, multilingual models like BharatGen's Param 2.
• The Evolving Landscape of LLM Evaluation — Recent AI advancements include top open-weights models, zero-shot MT, and evolving job market trends.
• The 1.6 Billion People Large Language Models Still Can’t Understand — Large language models fail profoundly in many non-English languages, going beyond mere performance degradation to confident incorrectness.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AIM Network.